Upload
cynthia-fields
View
219
Download
0
Embed Size (px)
Citation preview
Leng-Feng Lee Dec 3, 2004Slide 1 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Leng-Feng Lee ([email protected])
Advisor : Dr. Venkat N. Krovi
Mechanical and Aerospace Engineering Dept.
State University of New York at Buffalo
Decentralized Motion Planning within an Artificial Potential Framework (APF) for Cooperative
Payload Transport by Multi-Robot Collectives
Leng-Feng Lee Dec 3, 2004Slide 2 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Agenda
Motivation & System Modeling
Literature Survey & Research Issues
Part I
Part II
Local APF & limitations
Global APF-Navigation Function
Case Studies-Single robot with APF
Dynamic Formulation-Group of Robots
Motion Planning-Three Approaches
Case Studies-Multi Robots with APF
Performance Evaluation of Three Approaches
Conclusion & Future Work
Leng-Feng Lee Dec 3, 2004Slide 3 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
MotivationExamples of Multi-robot groups:
– Tasks are too complex;– Gain in overall performance;– Several simple, small-sized robot are easier, cheaper to built,
than a single large powerful robot system;– Overall system can be more robust and reliable.
Group Cooperation in Nature:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Schools of Fish Armies of Ants Flocks of Birds
How do we incorporate similar cooperation in artificial multi robot group?
Leng-Feng Lee Dec 3, 2004Slide 4 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Motivation
Example of Multi robot groups:
Cooperative payload transport
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Robots in formation
Robots in group
Leng-Feng Lee Dec 3, 2004Slide 5 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Why Motion Planning?
– To realize all the functionalities for mobile robots, the fundamental problem is getting a robot to move from one location to another without colliding with obstacles.
Motion Planning (MP) for Robot Collectives
Introduction Model Distribution Analysis Manipulability Cooperative System Conclusion
Definition:
– The process of selecting a motion and the associated set of input forces and torques from the set of all possible motions and inputs while ensuring that all constraints are satisfied.
MP for Robot Collective -
– MP exist for individual robots such as manipulator, wheeled mobile robot (WMR), car-like robot, etc.
– We want to examine extension of MP techniques to
Robot Collectives
Leng-Feng Lee Dec 3, 2004Slide 6 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Explicit Motion Planning:
– Decompose MP problem into 3 tasks:– Path Planning, Trajectory Planning, & Robot Control;
– Example: Road Map Method, Cell Decomposition, etc.
Motion Planning Algorithm Classification
Introduction Model Distribution Analysis Manipulability Cooperative System Conclusion
Implicit Motion Planning:
– Trajectory and actuators input are not explicitly compute before the motion occur.– Artificial Potential Field (APF) Approach belongs to this category.– Combine Path Planning, Trajectory Planning, and Robot Control in a single framework.
Leng-Feng Lee Dec 3, 2004Slide 7 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Motion Planning (cont’)
Introduction Model Distribution Analysis Manipulability Cooperative System Conclusion
Artificial Potential Field (APF) Approach:
– Obstacles generated a artificial Repulsive potential and goal generate an Attractive potential.
– Motion plan generated when attractive potential drives the robot to the goal and repulsive potential repels the robot away from obstacles.
– Combine Path Planning, Trajectory Planning, and Robot Control in one framework.
Subclass of Implicit Motion Planning Algorithm
Leng-Feng Lee Dec 3, 2004Slide 8 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Research Issues
Specific Research Questions:
– Which type of potential function is more suitable for MP for multi robot groups?
– How can we use the APF framework to help maintain formation? and
– How this framework be extended to realize the tight formation requirement for cooperative payload transport?
Introduction Model Distribution Analysis Manipulability Cooperative System Conclusion
Broad Challenges:
– Extending APF approach for Multi-robot collectives.
– Ensuring tight formations required for Cooperative Payload Transport application.
Leng-Feng Lee Dec 3, 2004Slide 9 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Research Issues (cont’)
Part I:
– Study various APF & their limitations;
– Determined a suitable APF as our test bed;
– Create a GUI to design and visualize the potential field;
– Case studies: MP for single robot using APF approach.
Introduction Model Distribution Analysis Manipulability Cooperative System Conclusion
To answer these research questions:
Part II:– E.O.M. for group of robots with formation constraints;– Solved the MP planning problem using three approaches; – Performance evaluation using various case studies.
Leng-Feng Lee Dec 3, 2004Slide 10 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Research Issues (cont’)
Introduction Model Distribution Analysis Manipulability Cooperative System Conclusion
Hierarchical difficulties in MP:
Our results:
– Multiple point-mass robots;
– Sphere World;
– Stationary Obstacles & Target.
(Dynamic Model)
Leng-Feng Lee Dec 3, 2004Slide 11 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
System ModelingIndividual level system models include:
– Point Mass Robot;– Differentially Driven Nonholonomic Wheel Mobile Robot (NH-WMR).
Introduction Model Distribution Analysis Manipulability Cooperative System Conclusion
,T
x yq , ,T
x y q
Leng-Feng Lee Dec 3, 2004Slide 12 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
System Modeling (cont’)
1 1 1 2 2 2 3 3 3, , , , , , , ,T
x y x y x y q
Introduction Model Distribution Analysis Manipulability Cooperative System Conclusion
Group level system model is formed using:
– Point Mass Robot;– Differentially Driven Nonholonomic Wheel Mobile Robot.
1 1 2 2 3 3, , , , ,T
x y x y x yq
Leng-Feng Lee Dec 3, 2004Slide 13 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
PART I: Artificial Potential Approach
Examine:
– Variants of APF & their limitations;
– Navigation function ;
– Single module formulations;
– Simulation studies.
Leng-Feng Lee Dec 3, 2004Slide 14 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF -background
Artificial Potential Field Approach– Proposed by Khatib in early 80’s.– FIRAS Function. [Khatib, 1986]
Later, various kind of Potential Functions were proposed:
– GPF Function. [Krogh, 1984]
– Harmonic Potential Function. [Kim, 1991]
– Superquadric Potential Function. [Khosla, 1988]
– Navigation Function. [Koditschek, 1988]
– Ge New Potential. [Ge, 2000]
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Leng-Feng Lee Dec 3, 2004Slide 15 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF Approach-Formulation
Idea:
– Goal generate an attractive potential well;
– Obstacle generate repulsive potential hill;
– Superimpose these two type of potentials give us the total potential of the workspace.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
) ( )( )( AtTot l pta ReUU U q q q
( )TotalU q denote the total artificial potential field;
( )AttU q denote the attractive potential field; and
( )RepU q is the repulsive potential field.
,T
i ix yq
Where:
is the position of the robot.
Leng-Feng Lee Dec 3, 2004Slide 16 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Target
Attractive Potential Field Gradient Plot
x-Position
y-P
ositio
n
Local APF -Attractive potential
Characteristics:
– Affect every point on the configuration space;
– Minimum at the goal.
– The gradient must be continuous.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Leng-Feng Lee Dec 3, 2004Slide 17 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF -Attractive potential
Example 1:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Where:
1( )
2
m
Att Tar RobU p q q
Tar Robq q = Euclidean distance between the robot and the target
Tarq
Robq
= Position of the target.
= Position of the robot.
= Positive scaling factor
2m is commonly used.
Leng-Feng Lee Dec 3, 2004Slide 18 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF -Attractive potential
Example 2:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Where:
For distance smaller than s, conical well. For distance larger than s, constant attractive force.
k = Positive scaling factor
2
2
, ( )
2 , Tar Rob Tar Rob
Att
Tar Rob Tar Rob
sU p
ks ks s
q q q q
q q q q
Leng-Feng Lee Dec 3, 2004Slide 20 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
-1.5 -1 -0.5 0 0.5 1 1.5-1.5
-1
-0.5
0
0.5
1
1.5
Obstacle
Repulsive Potential Field Gradient Plot
x-Position
y-P
ositi
on
Local APF -Repulsive potential
Characteristics:– The potential should have
spherical symmetry for large distance;
– The potential contours near the surface should follow the surface contour;
– The potential of an obstacle should have a limited range of influence;
– The potential and the gradient of the potential must be continuous.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Leng-Feng Lee Dec 3, 2004Slide 21 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF -Repulsive potential
Example 1 - FIRAS Function:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Where:
= Positive scaling factor
2
00
0
1 1 1, if
2
0 , if
RORep RO
RO
U
q
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
x position
y po
sitio
n
Contour plot of FIRAS Function of a Square Obstacle
Obstacle
1
2
3
4
5
6
7
8
9
RO Obs Rob q q
= the shortest Euclidean distance between the robot from the obstacle surface
RO
Leng-Feng Lee Dec 3, 2004Slide 22 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF -Repulsive potential
Example 2 - Superquadric Potential Function:– Approach Potential;– Avoidance Potential.– Avoid creation of local minima result from flat surface by creating
a symmetry contour around the obstacle.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Leng-Feng Lee Dec 3, 2004Slide 23 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF -Repulsive potential
Example 3 - Harmonic Potential Function:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
20 20
– Superimpose of another harmonic potential is also a harmonic potential. – More complicated shape can be modeled using ‘panel method’.
Repulsive Potential Attractive Potential
log2
r
Detail
Leng-Feng Lee Dec 3, 2004Slide 25 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF -Repulsive potential
Example 4 - Ge New Function:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Where:
–Modified from FIRAS function to solve the ‘Goal NonReachable for Obstacle Nearby’ -GNRON problem. – Ensures that the total potential will reach its global minimum, if and only if the robot reaches the target where
2
00
0
1 1 1, if
2
0 , if
nRT RO
Rep RO
RO
U
q
RT = Minimal Euclidean distance from robot to the target.
0RT 0 2 4 6 8 10 12 14 16 18 20
0
2
4
6
8
10
12
14
16
18
20
TargetObstacle
x position
y po
sitio
n
5
10
15
20
25
30
Obstacle
Leng-Feng Lee Dec 3, 2004Slide 26 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF -Repulsive potential
Potential Function with Velocities Information:
– Some potential function include the velocities information of the robots, obstacles and target.
– Example: Ge & Cui Potential [Dynamic obstacle & Target]. – Provide a APF for dynamic workspace.– Example: GPF Function. [Dynamic obstacles only].
– Can be used with our formulation for group of robots for motion planning in dynamic workspace.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Leng-Feng Lee Dec 3, 2004Slide 27 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF –Total Potential
Total Potential of Workspace:
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Tar
Target
Obstacle
Total Potential Field Gradient Plot
x-Position
y-P
ositi
on
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
( ) ( ) ( )Total Att Repi jU U U q q q
– Superimpose different repulsive potential from obstacles and different attractive potential from the goal, we get the total potential for the workspace.
– At any point of the workspace, the robot will reach the target by
following the negative gradient flow of the total potential.
( ) , , .Rep jU FIRAS Function Harmonic Function etcq
1( ) , .
2
m
Att Tar RobiU etc
q q q
Leng-Feng Lee Dec 3, 2004Slide 28 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF –Total Potential
Example: FIRAS Function
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
-3 -2 -1 0 1 2 3 4-3
-2
-1
0
1
2
3
4
x positiony
posi
tion
Contour plot of Potential Field Generated Using FIRAS Function
Obs
Obs
Target 5
10
15
20
0, 0 2, 2
1.5, -1.5
Rectangular Obstacle:
2 unit in height, 1 unit in width.
Circular Obstacle: Radius
Target :
0.5
More
Leng-Feng Lee Dec 3, 2004Slide 31 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF –Limitations
Local Minimum - result from single obstacle
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Target
Contour Plot of Attractive Potential Field
x position
y p
ositio
n
2
4
6
8
10
12
14
16
18
20
22
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
x position
y po
sitio
n
Contour plot of Repulsive Potential
Obstacle
1
2
3
4
5
6
7
8
9
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Local Minima
ObstacleTarget
x position
Contour plot Total Potential
y po
sitio
n
2
4
6
8
10
12
3D View
Leng-Feng Lee Dec 3, 2004Slide 33 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF –Limitations
Local Minimum - result from multiple obstacles
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Target
Obstacle
Local Minima
x position
y po
sitio
n
2
4
6
8
10
12
14
16
Leng-Feng Lee Dec 3, 2004Slide 34 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
-1 -0.5 0 0.5 1 1.5 2-1.5
-1
-0.5
0
0.5
1
1.5
Target(0.0, 0.0)
Influence Range ofRepulsive Potential
Obstacle
Goal Nonreachable with Obstacle Nearby, GNRON Problem
x position
y po
sitio
n
Local APF –Limitations
Limitation - Target close to obstacle:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
x position
y po
sitio
n
Contour Plot Showing GNRON Problem
Target
Local Minima(-0.4, 0.0)
Obstacle
1
2
3
4
5
6
7
8
9
10
11
Leng-Feng Lee Dec 3, 2004Slide 35 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Local APF -Limitations
Some other limitations include:
– No passage between closely spaced obstacle.– Non optimal path.– Implementation related limitations.
• Oscillation in the presence of obstacle;• Oscillation in narrow passages;• Infinite torque is not possible.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Leng-Feng Lee Dec 3, 2004Slide 36 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Global APF – Navigation Function
Properties:– Guarantee to provide a global minimum
at target.– Bounded maximum potential.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Let be a robot free configuration space, and let
Tarq be a goal point in the interior of , A map : [0,1] is a Navigation Function if it is:
1. Smooth on , that is, at least a 2 function.
2. Polar at Tarq ,i.e., has a unique minimum at Tarq on the path-connected
containing Tarq
.
component of
3. Admissible on , i.e., uniformly maximal on the boundary of
.
4. A Morse Function
[ Proposed by: Rimon & Koditschek]
Leng-Feng Lee Dec 3, 2004Slide 37 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Navigation Function
Navigation Function of a sphere world :
2 2i i i q q q
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Where:
2
1 1/2 for 0
( ) ( ) ( )
1 for 0
Tar
Total TarU
q q
q q q q q
2 20 0 0 q q q
0
M
ii
1,2i M Number of obstacles
is the implicit form of bounding sphere.
is the implicit form of obstacle geometric Eq.
Feature: Tunable by a single parameter :
Detail
Leng-Feng Lee Dec 3, 2004Slide 40 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Navigation Function
Example - Navigation Function of a sphere world :
0.4 4.0 Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
-6 -4 -2 0 2 4 6
-6
-4
-2
0
2
4
6
BoundedWorkspace
ConfigurationSpace
Obstacles
Target
A Euclidean Sphere World with 4 Obstacles
x Position
y P
ositi
on
Where:
Leng-Feng Lee Dec 3, 2004Slide 41 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Navigation Function -Constructions
At low value of , local minima may exist:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
-5 -4 -3 -2 -1 0 1 2 3 4 5
-5
-4
-3
-2
-1
0
1
2
3
4
5
x position
y po
sitio
n
Contour Plot of Navigation Function with = 2.6
Target
-5 -4 -3 -2 -1 0 1 2 3 4 5
-5
-4
-3
-2
-1
0
1
2
3
4
5
x position
y po
sitio
n
Contour Plot of Navigation Function with = 3.6
Target
3.6 2.6
Leng-Feng Lee Dec 3, 2004Slide 42 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Navigation Function – MATLAB GUI
A GUI to properly select a value:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Leng-Feng Lee Dec 3, 2004Slide 43 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
APF Approach – Formulation & Simulation
Idea:– We want the robot to follow the
negative gradient flow of the workspace potential field;
– Analogy to a ball rolling down to the lowest point in a given potential.
– Thus the gradient information will serve as the input to the robot system.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Leng-Feng Lee Dec 3, 2004Slide 44 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
APF Approach – Formulation
Formulation – Single point-mass robot:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Kinematic Model:
f Uqq u K
f UqMq u Kq K Kq
Dynamic Model:
Uq is the gradient of the potential field
2 2f fkK I is a positive diagonal scaling matrix
Kq is dissipative term added to stabilize the system
Leng-Feng Lee Dec 3, 2004Slide 45 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
APF Approach – FormulationFormulation – Nonholonomic Wheeled Mobile Robot (NH-WMR):
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Kinematic Model:
desired y-direction velocity. the desired x-direction velocity.
1 2
1
2
cos 0
sin 0
0 1
cos sin
,
p d d
d d
x
y u u
u k x y
u k atan2 x y
q
d
Ux
x
d
Uy
y
is the projected gradient onto the direction of forward velocity.
is the proportional to the angular error between the gradient and robot direction.
1u
2u
Leng-Feng Lee Dec 3, 2004Slide 46 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
APF Approach – Formulation
Formulation – Group robot without formation constraints:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Generalize position:
1 1 2 2, , , , ,T
n nx y x y x yq
Kinematic Model:
Dyanamic Model:
n -number of point-mass robot
f U qq u K
Mq u Kq f Uqu K
Leng-Feng Lee Dec 3, 2004Slide 47 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
APF Approach – Simulations
Simulation 1 – Single robot with single obstacle:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
f UqMq u Kq K Kq
2
00
0
1 1 1, if
2
0 , if
RORep RO
RO
U
qDetail
Leng-Feng Lee Dec 3, 2004Slide 49 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
APF Approach – SimulationsSimulation 2 – Single robot with two obstacles:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
f UqMq u Kq K Kq
2
00
0
1 1 1, if
2
0 , if
nRT RO
Rep RO
RO
U
qDetail
Leng-Feng Lee Dec 3, 2004Slide 53 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
APF Approach – Simulations
Simulation 3 – Single NH-WMR with four obstacles:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
2
1 1/2 for 0
( ) ( )
1 for 0
Tar
Total TarU
q q
q q q q
Detail
1 2
cos 0
sin 0
0 1
x
y u u
q
More
Leng-Feng Lee Dec 3, 2004Slide 55 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
APF Approach – Simulations
Simulation 4 – Group robots without formation constraint:
1( )
2
m
Att Tar RobU p q q
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
f UqMq u Kq K Kq
Detail
Leng-Feng Lee Dec 3, 2004Slide 57 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
APF Approach – Simulations
Simulation 5 – Group robots without formation constraint:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
1 1
q 0 I q 0
uq 0 M K q M
2
1/2 for 0
( ) ( )
1 for 0
Tar
Total TarU
q q
q q q qDetail
Leng-Feng Lee Dec 3, 2004Slide 59 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
PART II: Group Robots Dynamic Formulation
Include:– Dynamic Formulation for Group of Robots with Formation;– Solved the E.O.M using three Methods;
– Simulation Studies;
– Performance evaluation of each Methods.
Leng-Feng Lee Dec 3, 2004Slide 60 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Group Robots Dynamic Formulation
Approaches for formation maintenance:
– Formation Paradigm• Leader-follower [Desai et. al., 2001]
• Virtual structures [Lewis and Tan, 1997]
• Virtual leaders [Leonard and Fiorelli, 2001], [Lawton, Beard et al., 2003]
Our Approaches:
– View as a constrained mechanical system.
– Formation constraints – holonomic constraints added to a unconstrained dynamic system.
– Motion planning now can be treated as a forward dynamic simulation of a constrained mechanical system.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Leng-Feng Lee Dec 3, 2004Slide 61 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
The dynamic of group of robot can be formulated using Lagrange Equation by:
Group Robots Dynamic Formulation
q v
, ,T
t M q v f q, v u J q λ , t C q 0
q is the n-dimensional vector of generalized coordinates
v is the n-dimensional vector of generalized velocities
is the n-dimensional vector of generalized velocities M q
, , ,tf q v u is the n-dimensional vector of external forces
u is the vector of input forces, which is f U qk
C q
J q =q
is the Jacobian matrix
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
(1)
Leng-Feng Lee Dec 3, 2004Slide 62 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Group Robots Dynamic Formulation
The Lagrange Equation can be solved using following three methods:
– Method I: Direct Lagrange Multiplier Elimination Approach.• Explicitly computing the Lagrange multiplier by a projection into
the constrained force space.
– Method II: Penalty Formulation Approach.• Approximating the Lagrange multiplier using artificial compliance
elements such as virtual springs and dampers.
– Method III: Constraints Manifold Projection Based Approach• By projecting the equations of motion onto the tangent space of
the constraint manifold in a variety of ways to obtain constraint-reaction free equations of motions.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Leng-Feng Lee Dec 3, 2004Slide 63 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Group Robots Dynamic Formulation
2 1
12 1 , ,n
Tn t
vq
q M f q, v u J λ
Method I: Direct Lagrange Multiplier Elimination Approach:
– The direct Lagrange multiplier elimination is a centralized approach where the Lagrange multiplier is explicitly calculated to ensure formation constraints are not violated.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
(2)
(3)
The resulting Dynamic Equation can be expressed as:
2
1
2, ,t
t
-1 T -1 C
λ J q M J q J q M f q, v u J q q
Detail
Leng-Feng Lee Dec 3, 2004Slide 65 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Group Robots Dynamic Formulation
Method II: Penalty Formulation Approach:
– The holonomic constraints are relaxed and replaced by linear/non-linear spring with dampers.
– Here, the Lagrange multipliers are explicitly approximated as the force of a virtual spring or damper based on the extent of the constraint violation and assumed spring stiffness and damping constant.
Resulting Dynamic Equation:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
S D λ K C q K C qThis can be expressed as:
2 1
12 1 , , ,n
Tn S Dt
vq
q M f q v u J K C q K C q
(4)
Leng-Feng Lee Dec 3, 2004Slide 66 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Group Robots Dynamic Formulation
1, , ,t
T T Tυ S MS S f q v u S Mγ
2 1
1
2 1 , , ,
n
n m t
T T T
Sυ+ ηq
υ S MS S f q v u S Mγ
Method III: Constraints Manifold Projection Based Approach:
– In this approach, the dynamic equation with constraint-reactions is projected into the tangent space (feasible motion subspace) to obtain the constraint free projected dynamics equations.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
(5)
Thus, the resulting Dynamic Equation become:
Is the independent velocities. υ
Detail
Leng-Feng Lee Dec 3, 2004Slide 68 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Group Robots Dynamic Formulation
Baumgarte Stabilization:– To prevent numerical drift in the simulation, we adopted
Baumgarte stabilization method. – Baumgarte stabilization method involves the creation of an artificial
first or second order dynamical system which has the algebraic position-level constraint as its attractive equilibrium configuration.
For example, the holonomic constraint of Eq.(1) is replaced with:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
, 0 C q C q J q q C q 0
0tt e C C
Where the solution of the above equation is :
Leng-Feng Lee Dec 3, 2004Slide 69 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Case Study - Formulation:
Three point-mass robots forming a triangular shape:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
where:
The governing Equation can be written as:
T
q = v
M q q +V q,q +G q E q u J λ
J q q C q = 0
2 2
2 2
2 2
0 0
0 0
0 0
A
B
C
M
M q M
M
A
B
B
q
q q
q
V q,q Kq Tf U qu K
6 6E q I
G q = 0
, , , , ,T
A A B B C Cx y x y x yq
We will use this model to perform various case studies.
Leng-Feng Lee Dec 3, 2004Slide 70 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & Results
i j
Performance Evaluation – Formation Error:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Formation Error:
2 2 2
Error AB AB BC BC CA CAc c c c c c
Error is the total formation error;
ijc is the actual Euclidean distance between robot i and robot j
ijc is the desired Euclidean distance between robot and robot
Leng-Feng Lee Dec 3, 2004Slide 79 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulations & Results
Case Study 1 – Three robots in formation, without obstacle:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
1
T
vq
M u Kq J λq
Method I:
11 1
Tλ JM J JM u Kq J q q C q
Tf U qu K
Leng-Feng Lee Dec 3, 2004Slide 80 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulations & Results
Case Study 1 – Three robots in formation, without obstacle:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Method II:
2 1
12 1
nT
n S
vq
q M u Kq J K C q
1
1
1
, ,4 1
, ,4 1
, ,4 1
AAT
A A A A A A S A D A AA
BBT
B B B B B B S B D B BB
CCT
C C C C C C C S C D C C
vq
M E u K q J K C K Cq
vq
M E u K q J K C K Cq
vq
q M E u K q J K C K C
Decentralized Formulation:
Leng-Feng Lee Dec 3, 2004Slide 81 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & Results
Case Study 1 – Three robots in formation, without obstacle:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Method III:
1
T T T T
Sυ+ ηq
υ S MS S u S Mγ S Kq
1
2 2 2 2
1
2 2 2 2
1
2 2 2 2
A AA
TA A A A A A A
B BB
TB B B B B B B
B BC
TC C C C C C C
k
k
k
T
T
T
S υ ηq
υ S MS S I u M γ I q
S υ ηq
υ S MS S I u M γ I q
S υ ηq
υ S MS S I u M γ I q
Partial Decentralized Formulation:
Leng-Feng Lee Dec 3, 2004Slide 82 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & Results
0 1 2 3 4 5 6 7 8 9 100
1
2
3
4x 10
-14 Total Formation Error Using Method I
Tota
l F
orm
ation E
rror
Time, t0 1 2 3 4 5 6 7 8 9 10
0
0.005
0.01
0.015
0.02
0.025
0.03Total Error Using Method II
Time
Tota
l E
rror
Case Study 1 – Formation Error from three methods:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Method I
Method II
Method III
0 1 2 3 4 5 6 7 8 9 100
1
2
3
4
5
6
7x 10
-5 Total Formation Error Using Method III
Tota
l F
orm
ation E
rror
Time, t
Leng-Feng Lee Dec 3, 2004Slide 83 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & Results
Case Study 1 – Formation Error & Effect of Ks
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Method II:
0 1 2 3 4 5 6 7 8 9 1010
-7
10-6
10-5
10-4
10-3
10-2
10-1
Total Formation Error for Different Values of Ks using Method II
Time, t
Tota
l F
orm
ation E
rror
Ks=10
Ks=50
Ks=100
Ks=500
Ks=1000
Ks=5000
Ks=10000
Ks=10
Ks=50
Ks=100
Ks=500
Ks=1000
Ks=5000
Ks=10000
Leng-Feng Lee Dec 3, 2004Slide 84 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & Results
Case Study 1 – Formation Error & Effect of
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Method II:
0 1 2 3 4 5 6 7 8 9 10
10-10
10-9
10-8
10-7
Total Formation Error for Different Values using Method III
Time, t
Tota
l F
orm
ation E
rror
=10=50=100=500=1000
=10
=50
=100
Leng-Feng Lee Dec 3, 2004Slide 85 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & Results
-5 -4 -3 -2 -1 0 1 2 3 4 5-6
-5
-4
-3
-2
-1
0
1
2
3
4
5
Robots
RobotsFormations
Obstacle
Target
3 Robots carried a common payload - An 2D Workspace
x Position
y P
ositio
n
Case Study 2 – Three robots in Formation, one obstacle
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Leng-Feng Lee Dec 3, 2004Slide 86 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulations & Results
Case Study 2 – Three robots in formation, one obstacle:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Method II Method III Method I
Leng-Feng Lee Dec 3, 2004Slide 88 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & Results
Case Study 2 – Formation Error from three methods:
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Method I
Method II
Method III
Leng-Feng Lee Dec 3, 2004Slide 89 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & Results
0 1 2 3 4 5 6 7 810
-5
10-4
10-3
10-2
10-1
100
Time, t
Tot
al E
rror
Total Formation Error For Different value of
=10=100=300=500=700=1000
=100
=300
=500 =700
=1000
=10
Case Study 2 – Formation Error & Effect of Ks &
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Method II Method III
0 1 2 3 4 5 6 7 80
1
2
3
4
5
6x 10
-7
Time, t
Tot
al E
rror
Total Formation Error For Different value of
=10=20=30=40=50=60=70
=70
=60
=50
=40
=30
Leng-Feng Lee Dec 3, 2004Slide 90 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & Results
Case Study 2 – Three robots in Formation with Expansion.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
2 0.5 , 0 4
4, 4ij
t tc
t
Each sides change from 2 units to 4 units in 4 seconds:
Leng-Feng Lee Dec 3, 2004Slide 91 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & Results
Case Study 3 – Three robots in Formation with Expansion.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Method II Method III Method I
Leng-Feng Lee Dec 3, 2004Slide 93 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & Results
Case Study 3 – Formation Error & Effect of Ks &
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Method II
Method III
Leng-Feng Lee Dec 3, 2004Slide 94 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & ResultsCase Study 4 – Three robots in Formation with Shape Change.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Constraint between robot A & B change from 2 units to 4 units in 4 seconds:
2 0.5 , 0 4
4, 4AB
t tc
t
Note: Method I cannot perform this task because when three robots in a straight line, the inverse of the Jacobian matrix become singular.
Leng-Feng Lee Dec 3, 2004Slide 95 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & Results
Case Study 4 – Three robots in Formation with Shape Change.
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Method II Method III
Leng-Feng Lee Dec 3, 2004Slide 96 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Simulation & ResultsCase Study 4 – Formation Error & Effect of Ks &
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
Method II: Method III:
Leng-Feng Lee Dec 3, 2004Slide 97 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Conclusion
General Characteristics – Formation Accuracy
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
The average total formation error for each method :
Method III Method I Method IIError Error Error
Leng-Feng Lee Dec 3, 2004Slide 98 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Conclusion
Method I Method II Method IIITime Time Time
General Characteristics – Computational Time
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
The average total Computational Time (sec) for each method :
Leng-Feng Lee Dec 3, 2004Slide 99 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Conclusion
Method I Method III
General Characteristics – Decentralize formulation capability
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
The decentralize formulation capability for each method :
Centralized Decentralized
Method II
Leng-Feng Lee Dec 3, 2004Slide 100 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Conclusion
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
General Characteristics – Formation related concerns:
– The Jacobian matrix in Method I and Method III can become singular in some specific position.
– Method II has no such limitations.
In Summary:
Leng-Feng Lee Dec 3, 2004Slide 101 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Conclusion
Introduction APF Approaches Group Robot Formulation Simulation Results Conclusion
– Evaluation of various potential functions.
– Development of a GUI to generate navigation function.
– Develop the group motion planning problem as a forward dynamic simulation problem;
– Evaluation of three different method in solving motion planning problem for a group of robots in formation.
– Critical evaluation of the performance by the three approaches.
Leng-Feng Lee Dec 3, 2004Slide 102 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Future Work
Introduction Model Distribution Analysis Manipulability Cooperative System Conclusion
– Provide a way to avoid Jacobian matrix become singular.
– Incorporate nonholonomic constraints in the formulation.
– Implement a more efficient gradient finding method by utilizing the available information from each robot.
– Implement the algorithm in a decentralized computation manner.
Leng-Feng Lee Dec 3, 2004Slide 103 of 51 Automation, Robotics and Mechatronics Lab, SUNY at Buffalo
Thank You!Acknowledgments:
Dr. V. Krovi, Dr. T. Singh & Dr. J. L. Crassidis