Transcript AiLabSeminar_BulucCelik

AI Lab
Weekly Seminar
By: Buluç Çelik
25.03.2005
1
General Outline
► Part
I: A Behavior Architecture for Autonomous
Mobile Robots Based on Potential Fields
► Part
II: Real-Time Object Recognition Using
Decision Tree Learning
► Part
III: My Thesis - Comparison of MultiAgent Planning Algorithms
25.03.2005
2
Part I
A Behavior Architecture for Autonomous
Mobile Robots Based on Potential Fields
Laue, T., Röfer, T. (2005)
In: 8th International Workshop on RoboCup 2004
(Robot World Cup Soccer Games and Conferences),
Lecture Notes in Artificial Intelligence. Springer, im
Erscheinen.
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
3
Outline
► 1.
Introduction
► 2.
Architecture
► 3.
Modeling of the Environment
► 4.
Motion Behaviors
► 5.
Behaviors for Action Evaluation
► 6.
Applications
► 7.
Conclusion & Future Works
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
4
1. Introduction
► Artificial
Potential Fields
 Popular for being capable of acting in
continuous domains in real time
 Can follow a collision-free path via the
computation of a motion vector from the
superposed force fields
►Repulsive
force fields to obstacles
►Attractive force fields to desired destination
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
5
1. Introduction
► Behavior-based
Architectures
 The proposed approach combines existing
approaches in a behavior based architecture by
realizing single competing behaviors as potential
fields
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
6
2. Architecture
► Potential
fields are based on superposion of
force fields
 Fails for tasks with more than one possible goal
position (e.g. goalkeeper)
 Could be solved by selecting the most
appropriate goal
►But
this proceeding will affect the claim of standalone architecture
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
7
2. Architecture
► Different
tasks have to be splitted into different
competing behaviors
► Among the blocking and keeping of behaviors
under certain circumstances, behaviors can be
combined with others to realize small hierarchies
 For instance, this allows the usage of a number of
evaluation behaviors differentiating situations (e. g.
defense or midfield play in robot soccer) respectively
combined with appropriate motion behaviors
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
8
3. Modeling of the Environment
► The
architecture offers various options
allowing a detailed description
 An object class O: O = (fO, GO, FO)
►fO :
potential function (e.g. attractive, repulsive)
►GO : geometric primitive used to approximate an
object’s shape
►FO : the kind of field (e.g. circumfluent around GO ,
tangential around the position of the instance)
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
9
4. Motion Behaviors
► The
general procedure of motion planning
 A vector v can be computed by the
superposition of the force vectors vi of all n
object instances assigned to a behavior
n
v   vi R 
i 1
is the current position of the robot
►v can be used to determine the robot’s direction of
motion, rotation and speed
►R
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
10
4. Motion Behaviors
► Relative
motions
 Assigning force fields to single objects of the
environment allows the avoidance of obstacles and the
approach to desired goal positions
 Moving to more complex spatial configurations (e. g.
positioning between the ball and the penalty area or
lining up with several robots) is not possible directly
 Relative motions are realized via special objects which
may be assigned to behaviors
► Such
an object consists of a set of references to object
instances and a spatial relation (e. g. between, relative-angle)
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
11
4. Motion Behaviors
► Dealing
with local minima
 Local minima are an inherent problem of potential
fields, an optimal standard solution does not exist
► The
attractive potential
guides the robot’s path into a
C-obstacle concavity
► At some point, the repulsive
force cancels exactly the
attractive force
►This stable zero-force
configuration is a local minimum of the total potential function
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
12
4. Motion Behaviors
► Dealing
with local minima
 A* algorithm is used, the continuous environment is
discretized
 A search tree
with a dynamic
number and
size of
branches is
build up
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
13
5. Behaviors for Action Evaluation
► An
action behavior can be combined with a motion
behavior to determine the appropriateness of its
execution
► The environment is rasterized into cells of fixed size
► The anticipated world state after an action is
computed
► The value of only the relevant positions are
evaluated to determine the most appropriate
position
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
14
5. Behaviors for Action Evaluation
may be determined at an arbitrary position P,
being the sum of the potential functions of all object
instances assigned to the behavior
► φ(P)
n
 P     i P 
i 1
► evaluate
a certain action which changes the
environment (e. g. kicking a ball) this action has to
be mapped to a geometric transformation (e. g.
rotation, translation) in order to describe the motion
of the manipulated object
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
15
6. Applications
►
The architecture has been applied to two different
platforms, both being RoboCup teams of the
Universit¨at Bremen
 Robots of the Bremen Byters, which are a part of the
GermanTeam (Sony Four-legged Robot League)
 The control program of B-Smart, which competes in the
RoboCup F-180 (Small Size) league
For playing soccer, about 10–15 behaviors have been
needed (e. g. go to Ball, go to defense position or kick
ball forward)
► Sequences of actions have been used, allowing a quite
forward-looking play
►
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
16
7. Conclusion & Future Works
►
A behavior-based architecture
 For autonomous mobile robots
 Integrating several different approaches for motion planning and
action evaluation into a single general framework
 Dividing different tasks into competing behaviors
►
Future works
 Porting the architecture to other platforms to test and extend the
capabilities of this approach
 There exist several features already implemented but not adequately
tested (e. g. the integration of object instances based on a
probabilistic world model)
 In addition, the behavior selection process is currently extended to
deal with a hierarchy of sets of competing behaviors, allowing the
specification of even more complex overall behaviors
25.03.2005
A Behavior Architecture for Autonomous Mobile Robots Based on Potential Fields
17
Part II
Real-Time Object Recognition Using Decision
Tree Learning
Wilking, D., Röfer, T. (2005)
In: 8th International Workshop on RoboCup 2004
(Robot World Cup Soccer Games and Conferences),
Lecture Notes in Artificial Intelligence. Springer, im
Erscheinen.
25.03.2005
Real-Time Object Recognition Using Decision Tree Learning
18
Outline
► 1.
Introduction
► 2.
The Recognition Process
► 3.
Results
► 4.
Conclusion
25.03.2005
Real-Time Object Recognition Using Decision Tree Learning
19
1. Introduction
► The
goal of the process presented in this paper is
the computation of the pose of a visible robot
(i. e. the distance, angle, and orientation)
► Apart from the unique color which can be used
easily to find a robot in an image, the geometric
shapes of the different parts provide much more
information about the position of the robot
► The shapes themselves can be approximated using
simple line segments and the angles between
them
25.03.2005
Real-Time Object Recognition Using Decision Tree Learning
20
2. The Recognition Process
The recognition begins with iterating through the
surfaces that have been discovered
by the preprocessing stage
► For every surface, a number
of segments approximating
its shape and a symbol is
generated (e. g. head, side,
front, back, leg, or nothing)
► The symbols are inserted into
a special 180 symbol memory
►
25.03.2005
Real-Time Object Recognition Using Decision Tree Learning
21
2. The Recognition Process
► Segmentation
and surface detection
 Relevant pixels are determined by color
segmentation using color tables
 Surfaces (along with their position,
bounding box, and area) are computed
 The contour of the surface is computed
 The iterative-end point algorithm is used
to compute the segments
25.03.2005
Real-Time Object Recognition Using Decision Tree Learning
22
2. The Recognition Process
► Attribute
generation
 Simple attributes (e. g., color class, area, perimeter,
and aspect ratio)
 Regarding the representation of the surface (e. g.
line segments, the number of corners, the
convexity and the number of different classes of
angles between two line segments)
 The surface is compared to a circle and a rectangle
with the same area
 Sequences of adjacent angles
25.03.2005
Real-Time Object Recognition Using Decision Tree Learning
23
2. The Recognition Process
► Classification
 The decision tree learning algorithm is chosen
as classification algorithm
 The tree is built by calculating the attribute with
the highest entropy
 over-fitting is solved using χ2-pruning
25.03.2005
Real-Time Object Recognition Using Decision Tree Learning
24
2. The Recognition Process
► Analysis
 The surface area of a group is used
to determine the distance to the
robot
 The direction to the robot is
computed by the group’s position in
the 180 memory
 The relative position of the head
within the group and the existence
of front or back symbols indicate the
rough direction of the robot
25.03.2005
Real-Time Object Recognition Using Decision Tree Learning
25
3. Results
25.03.2005
Real-Time Object Recognition Using Decision Tree Learning
26
3. Results
25.03.2005
Real-Time Object Recognition Using Decision Tree Learning
27
4. Conclusion
►A
robot recognition process based on decision tree
classification
► Due to the complexity and length of the process,
some parts could be streamlined
► The heuristics used during the analysis step can
be improved using a skeleton template based,
probabilistic matching procedure
 deal both with the problem of occlusion and missing
symbols
► improvements
concerning the speed of the
attribute generation can be achieved
25.03.2005
Real-Time Object Recognition Using Decision Tree Learning
28
Part III
My Thesis
Comparison of Multi-Agent Planning
Algorithms
25.03.2005
Comparison of Multi-Agent Planning Algorithms
29
Comparison of Multi-Agent
Planning Algorithms
► Multi-agent
planning algorithms are to be
designed and implemented for Sony Four-legged
Robot League
► A behavior architecture for autonomous mobile
robots based on potential fields will be designed
and implemented
 One similar to the one explained at Part I
► Training
a neuro-fuzzy system using the designed
behavior architecture
 A neuro-fuzzy system will be trained using the decisions
made by the implementation of behavior architecture
25.03.2005
Comparison of Multi-Agent Planning Algorithms
30
Comparison of Multi-Agent
Planning Algorithms
► Training
the neuro-fuzzy system with
playing against the behavior architecture
 The neuro-fuzzy system will be trained more by
playing against the implementation of behavior
architecture
►A
decision tree will be produced from the
neuro-fuzzy network
► The architectures will be evaluated and
compared
25.03.2005
Comparison of Multi-Agent Planning Algorithms
31
Thank You...
Discussion
25.03.2005
32