Swarm Intelligence in Robotics

Swarm Intelligence in Robotics

Dr. Alaa Khamis Pattern Analysis and Machine Intelligence University of Waterloo [email protected] MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

1/22 1

Outline • Introduction • Body/Brain Evolution

MUSES_SECRET: ORF-RE Project - © PAMI Research Group – University of Waterloo

• Traditional Robotics • Swarm Robotics • Swarm Intelligence (SI) • SI in Traditional Robotics • SI in Swarm Robotics • Concluding Remarks

MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

2/22 2




3/22 3


Introduction • PAMI Research Group Synchromedia Synchromedia Lab Lab

CIM CIM Lab Lab

Intelligent Multi-crane CRS F3 and CRS T265

Pattern Pattern Analysis Analysis Lab Lab Speech Understanding

Pattern Analysis

Mobile Mobile Robotics Robotics Lab Lab Magellan Pro

Computer vision

Data Mining Services Catalog


ATRV-mini Soccer-playing Robots

4/22 4

Introduction • Swarm Intelligence in Robotics The objective of this talk is to highlight the different applications of the rapidly emerging field of swarm intelligence in solving complex problems of traditional and swarm robotics.


5/22 5





6/22 6

Body/Brain Evolution Brains SI

Swarm Robotics

Cognitive Robotics

100s

DAI

Multiagent

Distributed Robot System

10s

AI 1

Robotics

Centralized Control

Machine

Multiple Machines

1

10s


MEMS-based Multiple Machines Bodies 100s 7/22 7





8/22 8

Traditional Robotics The Evolutionary Stages

Industrial Robotics

Service Robots for Professional Use

Service Robotics Evolution


Service Robots for Personal Use

Personal Robotics

9/22 9

Traditional Robotics • Traditional Problems Environment Perception SLAM Path Planning Navigation Autonomy Human Interaction Learning


10/22 10




11/22 11


Swarm Robotics • Definition Swarm robotics is the study of how large number of relatively simple physically embodied agents can be designed such that a desired collective behavior emerges from the local interactions among agents and between the agents and the environment. Resolving complexity. Increasing performance. Simplicity in design. Reliable. Box Pushing


12/22 12

Swarm Robotics • Typical Problem Domains Box-pushing Foraging Aggregation and Segregation Formation Control Cooperative Mapping Soccer Tournaments Site Preparation Sorting

http://www.robocup.org/

Collective Construction

http://www.fira.net/


13/22 13

Swarm Robotics • Applications of Swarm Robotics Autonomous inspection of complex engineered structures. Distributed sensing tasks in micromachinery or the human body. Killing Cancer Tumors in Human Body Mining Agricultural Foraging Cooperative Tracking UAV drones Interactive Art Space-based Construction Spy robots Rescue Operations Humanitarian Demining Surveillance, Reconnaissance and Intelligence. Mobile Trackers MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

14/22 14

Swarm Robotics • Projects

http://www.swarm-bots.org/

The SWARM-BOT project aims to study a novel swarm robotics system. It is directly inspired by the collective behavior of social insects and other animal societies. It focuses on self-organization and self-assembling of autonomous agents. Its main scientific challenge lays in the development of a novel hardware and of innovative control solutions.

Swarm-bots, Marco Dorigo, 2005 15/22 15


Swarm Robotics • Projects

PAMI Lab http://horizon.uwaterloo.ca/multirobot/

Leaderless Distributed (LD) Flocking Algorithm Flocking in Embedded Robotic Systems CORO - Caltech


16/22 16

Swarm Robotics • Team Size and Composition Team Size

Alone

Pair

Team Composition

Limited Group

Infinite Group

Homogenous

Heterogeneous

17/22 17


Swarm Robotics • Team Reconfigurability Team Reconfigurability

Static

coordinated


Dynamic

18/22 18

Swarm Robotics • Communication Pattern Implicit Communication

Explicit Communication

Broadcast

Graph

Interaction via Environment (Stigmergy)

Interaction via Sensing

Action Action

Perception

Environment

Perception

Virtual Pheromone

Perception

1 to

6

Address or Directed Messages

11 to 8

19/22 19


Swarm Robotics • Communication Range and Bandwidth Communication Bandwidth

Communication Range

None

Near

Infinite

High


Moderate

Low

zero

20/22 20

Swarm Robotics • Challenging Problems Coordination Algorithm Design Implementation and Test Analysis and Modelling

21/22 21


Swarm Robotics • Challenging Problems: Coordination Coordination

Cooperation

Collaboration

Planning

Centralized

Decentralized

Hierarchical

Competition

Negotiation

Holonic


22/22 22

Swarm Robotics • Challenging Problems: Algorithm Design Swarm roboticists face the problem of designing both the physical morphology and behaviours of the individual robots such that when those robots interact with each other and their environment, the desired overall collective behaviours will emerge. At present there are no principled approaches to the design of low-level behaviours for a given desired collective behaviour [1]. “collective behavior is not simply the sum of each participant’s behavior, as others emerge at the society level” [2]. MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

23/22 23

Swarm Robotics • Challenging Problems: Implementation and Test To build and rigorously test a swarm of robots in the laboratory requires a considerable experimental infrastructure. Real-robot experiments thus typically proceed hand-in-hand with simulation and good tools are essential [1].


24/22 24

Swarm Robotics • Challenging Problems: Implementation and Test Advanced Robotics Interface for Applications (ARIA): Robotic Sensing and Control Libraries. Open Robot Control Software (OROCOS): open-source real time control architecture for different machines. Microsoft Robotics Studio: is a Windows-based environment for robot control and simulation. Player/Stage/Gazebo: PSG is open source software that used and developed by an international community of researchers from over 30 universities/companies. 25/22 25


Swarm Robotics • Challenging Problems: Implementation and Test Stage 3.0 [3]

2,000 minibots MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

100 Pioneer Robots 26/22 26

Swarm Robotics • Challenging Problems: Analysis and Modelling A robotic swarm is typically a stochastic, non-linear system and constructing mathematical models for both validation and parameter optimization is challenging. Such models would surely be an essential part of constructing a safety argument for real-world applications [1].


27/22 27





28/22 28

Swarm Intelligence DNA

Cellular Automata

Brain

Neural Networks Protection of computers and networks

Immune System Evolution

Genetic Algorithms

Social Insects

Swarm Intelligence


29/22 29

Swarm Intelligence Analogies

Nature

Unmanned Autonomous Vehicles (UAVs)

Ants/UAVs

Prey/Target

Predators/Threats

Environment/ Battlefield


30/22 30

Swarm Intelligence • Definition Swarm intelligence (SI) refers to the phenomenon of a system of spatially distributed individuals coordinating their actions in a decentralized and self-organized manner so as to exhibit complex collective behavior. “SI is the property of a system whereby the collective behaviors of (unsophisticated) agents interacting locally with their environment cause coherent functional global patterns to emerge.” Swarm intelligence solves optimization problems. 31/22 31


Swarm Intelligence • Key Elements A large number of “simple” processing elements; Neighborhood communication; Though convergence in guaranteed, time to convergence is uncertain. Most of the research are experimental: Observation

Create Create

Extract Extract

Metaheuristic Build Build

Test Test

Refine

Model

Simulation Algorithm MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

32/22 32

Swarm Intelligence • SI-based Approaches Initialize parameters Initialize population While (end condition not satisfied ) loop over all individuals Find best so far Find best neighbor Update individual


33/22 33

Swarm Intelligence • SI-based Approaches Ant Colony Optimization (ACO) Particle Swarm Optimization (PSO) Stochastic Diffusion Search (SDS)


34/22 34





35/22 35

SI in Traditional Robotics • ACO-based Motion Planning [4] • PSO-based Motion Planning [5] • Wall-following autonomous robot (WFAR) navigation [6] • Gait Optimization [7] • Kinematics and Dynamics of Robot Manipulators [8,9] • Learning [10]


36/22 36

SI in Traditional Robotics • Motion Planning Path planning is a process to compute a collision-free path for a robot from a start position to a given goal position, amidst a collection of obstacles. T

S MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

37/22 37

SI in Traditional Robotics • ACO-based Motion Planning [4] Proposed Method Step 1: MAKLINK graph theory to establish the free space model of the mobile robot; Step 2: utilizing the Dijkstra algorithm to find a suboptimal collision-free path; Step 3: utilizing the ACS algorithm to optimize the location of the sub-optimal path so as to generate the globally optimal path.


38/22 38

SI in Traditional Robotics • ACO-based Motion Planning [4] Step 1: MAKLINK-based Free Space Model 1. The heights of the environment and obstacles can be ignored; 2. There exist some known obstacles distributed in the environment, both the environment and the obstacles have a polygonal shape; 3. In order to avoid a moving path too close to the obstacles, the boundaries of every obstacle can be expanded.


39/22 39

SI in Traditional Robotics • ACO-based Motion Planning [4] Step 1: MAKLINK-based Free Space Model Real Obstacle

Grown Obstacle

Free MAKLINK line: 1) Either its two end points are two vertices on two different grown obstacles or one point is a vertex of a grown obstacle and the other is located on a boundary of the environment; 2) Every free MAKLINK line cannot intersect any of the grown obstacles. MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

boundary 40/22 40

SI in Traditional Robotics • ACO-based Motion Planning [4] Step 1: MAKLINK-based Free Space Model [11,5] ◊ Constructing MAKLINK graph 1. Find all the lines that connect one of the corners that belong to a polygonal obstacle, with all the other obstacles’ corners including the corners of the current obstacle; S

T 41/22 41


SI in Traditional Robotics • ACO-based Motion Planning [4] Step 1: MAKLINK-based Free Space Model [11,5] ◊ Constructing MAKLINK graph 2. Delete the redundant free links to make every free space, of which the edges are free links, obstacle edges and boundary walls, be a convex polygon and its area be largest; S

T MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

42/22 42

SI in Traditional Robotics • ACO-based Motion Planning [4] Step 1: MAKLINK-based Free Space Model [11,5] ◊ Constructing MAKLINK graph 3. Find the midpoint of the remained free links and take them as the path nodes, labeling orderly as 1, 2,…, n. The connections among the midpoints that belong to the same convex area compose a network. S

T MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

43/22 43

SI in Traditional Robotics • ACO-based Motion Planning [4] Step 1: MAKLINK-based Free Space Model v1, v2, . . . , vl are the middle points of these free MAKLINK lines; l is total number of the free MAKLINK lines on a MAKLINK graph. l=26 v0 is S vl+1 is T

Network graph for free motion of robot


44/22 44

SI in Traditional Robotics • ACO-based Motion Planning [4] Step 2: Dijkstra Algorithm Sub-optimal path S →v1 →v2 →v10 →v11 →v12 →v13 → v16 →v17 →v24 →T

Path length= 507.692 m. This path is just a sub-optimal path because it passes only through the middle points of those free MAKLINK lines Sub-optimal path generated by Dijkstra 45/22 45


SI in Traditional Robotics • ACO-based Motion Planning [4] Step 3: ACS Algorithm Assume sub-optimal path generated by the Dijkstra algorithm is

P0 → P1 → P2 → ... → Pd +1 These path nodes lie on the middle points of the relevant free MAKLINK lines. Now, we need to adjust and optimize their locations on their corresponding free MAKLINK lines.

Pi = Pi1 + ( Pi 2 − Pi1 ) × hi , Pi = Pi1

i = 1, 2 ,..., d

hi = 0

Pi = ( Pi1 + Pi 2 ) / 2 Pi = Pi 2

hi ∈ [ 0 ,1],

hi = 0 . 5 midpoint

hi = 1


Path Coding Method

46/22 46

SI in Traditional Robotics • ACO-based Motion Planning [4] Step 3: ACS Algorithm The objective function of the optimization problem will be:

L=

d

∑ length { P ( h ),P i=0

i

i

i +1

( hi +1 )}

ACS algorithm is used to find the optimal parameter set

{ h *1 , h * 2 ,..., h * d } such that L has the minimum value.


47/22 47

SI in Traditional Robotics • ACO-based Motion Planning [4] Step 3: ACS Algorithm Grid graph on which there are dx11 nodes in total. nij is node j on line hi. The path of an ant depart from the starting point S is

S → n1 j → n 2 j → ... → n dj → T Generating of nodes and moving paths MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

48/22 48

SI in Traditional Robotics • ACO-based Motion Planning [4] Step 3: ACS Algorithm ◊ Pheromone Concentration

Assume that at initial time t =0

All the nodes have the same pheromone concentration τ0:

τ ij ( 0 ) = τ o ( i = 1, 2 ,..., d ;

j = 0 ,1, 2 ,..., 10 )

49/22 49


SI in Traditional Robotics • ACO-based Motion Planning [4] Step 3: ACS Algorithm ◊ Transition Rule

Assume the number of ants is m.

In moving process, for an ant k, when it locates on line hi-1, it will choose a node j from the eleven nodes of the next line hi to move to according to the following rule:

 arg max j= J ,

u∈ A

{[τ iu ( t )].[ η iu ] β },


if q ≤ q o

if q f q o 50/22 50



u∈ A

{[τ iu ( t )].[ η iu ] β },

if q ≤ q o

if q f q o

where A represents the set: {0, 1, 2, . . . , 10}; τiu(t) is the pheromone concentration of node niu;

51/22 51




β τ η {[ ( t )].[ ] }, u∈ A iu iu

if q ≤ q o

if q f q o

ηiu represents the visibility of node niu and is computed by the following equation: *

η ij =

1 . 1 − y ij − y

ij

1 .1

yij is the y-coordinate of node nij and y*ij are values that are corresponding to the path nodes on the optimal robot path generated by the ACS algorithm in the previous iteration. MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

52/22 52



u∈ A

{[τ iu ( t )].[ η iu ] β },

if q ≤ q o

if q f q o

β is an adjustable parameter which controls the relative importance of visibility ηiu versus pheromone concentration τiu(t); q is a random variable uniformly distributed over [0, 1]; q0 is a tunable parameter (0≤q0≤ 1); 53/22 53




β τ η {[ ( t )].[ ] }, u∈ A iu iu

if q ≤ q o

if q f q o

J is a node which is selected according to the following probability formula and “Roulette Wheel Selection Method”

P (t ) = k iJ

[τ iJ ( t )].[ η iJ ] β 10

[τ ∑ ω =0

β ( t )].[ η ] iw iw


54/22 54

SI in Traditional Robotics • ACO-based Motion Planning [4] Step 3: ACS Algorithm ◊ Pheromone Update: After Each Iteration

τ ij ( t ) ← (1 − ρ ).τ ij ( t ) + ρ .∆ τ ij ( t ) 0 p ρ p1

∆ τ ij ( t ) =

1 L+

where L+ is the length of robot path corresponding to the best ant tour. MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

55/22 55

SI in Traditional Robotics • ACO-based Motion Planning [4] Step 3: ACS Algorithm ◊ Pheromone Update: After passing through a node nij

τ ij ( t ) ← (1 − ρ ).τ ij ( t ) + ρ .τ o When a node is visited several times by ants, repeatedly applying the local updating rule will make the pheromone level of this node diminish. This has the effect of making the visited nodes less and less attractive to the ants, which indirectly favors the exploration of not yet visited nodes. MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

56/22 56

SI in Traditional Robotics • ACO-based Motion Planning [4] Results ACS Algorithm

Realcoded GA

Average CPU time per iteration (sec.)

0. 00059

0. 00067

Average number of iterations needed for convergence

175

912

Average CPU time needed for obtaining optimal solution (sec.)

0.1033

0.6110

57/22 57


SI in Traditional Robotics • PSO-based Motion Planning [5] Sub-optimal path from Dijkstra algorithm is

P0 → P1 → P2 → ... → Pd +1 These path nodes lie on the middle points of the relevant free MAKLINK lines. Location adjustment:

Pi = Pi1 + ( Pi 2 − Pi1 ) × hi , Pi = Pi1

i = 1, 2 ,..., d

hi = 0

Pi = ( Pi1 + Pi 2 ) / 2 Pi = Pi 2

hi ∈ [ 0 ,1],

hi = 0 . 5 midpoint

hi = 1 Path Coding Method


58/22 58

SI in Traditional Robotics • PSO-based Motion Planning [5] Each particle Xi is constructed as:

X i = ( h1 , h2 ,... h d )

Pi ( hi )

Euclidian Distance

The fitness value of each particle is:

f (Xi) =

d


i

i +1

i

( hi +1 )}

Pi +1 ( hi +1 )

The smaller the fitness value is, the better the solution is.

59/22 59


SI in Traditional Robotics • PSO-based Motion Planning [5] 1. Initialize particles at random, and set pBest= Xi; 2. Calculate each particle’s fitness value according to equation

f (Xi) =

d

∑ length { P ( h ), P i=0

i

i

i +1

( hi + 1 )}

and label the particle with the minimum fitness value as gBest;


60/22 60

SI in Traditional Robotics • PSO-based Motion Planning [5] 3. For k1=1 to k1max do { 4. For each particle Xi do { 5. Update vi and xi according to the following equations: vi ( k + 1) = ω × vi ( k ) + c1 × rand () × [ pBest i ( k ) − xi ( k )] + c 2 × rand () × [ gBest ( k ) − xi ( k )] xi ( k + 1) = xi ( k ) + vi ( k + 1),

1 ≤ i ≤ n,

6. Calculate the fitness according to equation f (Xi) =

d


i

i

i +1

( hi +1 )}

7. Update gBest and pBesti ; 8. If ||v|| ≤ε , terminate ;} MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

61/22 61

SI in Traditional Robotics • PSO-based Motion Planning [5]


62/22 62





63/22 63

SI in Swarm Robotics • Multirobot Control [12] • Collective Robotic Search (CRS) [13] • Odor-source Localization [14,15] • Mobile Sensor Deployment [16] • Learning [17] • Communication Relay [18] • Group Behavior [19]


64/22 64

SI in Swarm Robotics • Multirobot Control [12] Unintelligent carts are commonly found in large airports. Travelers pick up carts at designated points and leave them in arbitrary places. It is a considerable task to re-collect them. It is, therefore, desirable that intelligent carts (intelligent robots) draw themselves together automatically.


65/22 65

SI in Swarm Robotics • Multirobot Control [12] Objectives Using mobile software agents to locate robots scattered in a field, e.g. an airport, and make them autonomously determine their moving behaviors by using an ACO-based clustering algorithm.


66/22 66

SI in Swarm Robotics • Multirobot Control [12] Assumptions - Each robot has a function for collision avoidance; - Each robot has a function to sense RFID embedded in the floor carpet to detect its precise coordinate in the field.

A team of mobile robots work under control of mobile agents

RFID under a carpet tile 67/22 67


SI in Swarm Robotics • Multirobot Control [12] Quasi Optimal Robots Collection Step 1: One mobile agent issued from the host computer visits scattered robots one by one and collects the positions of them.

Mobile Agent

Host Computer

Assembly Point

Assembly Point Intelligent Cart

Assembly Point MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

68/22 68

SI in Swarm Robotics • Multirobot Control [12] Quasi Optimal Robots Collection Step 2: Another agent called the simulation agent, performs ACC algorithm and produces the quasi optimal gathering positions for the mobile robots. The simulation agent is a static agent that resides in the host computer. Assembly Point

Host Computer


Simulation Agent ACO-based Clustering

Assembly Point 69/22 69


SI in Swarm Robotics • Multirobot Control [12] Quasi Optimal Robots Collection Step 3: A number of mobile agents are issued from the host computer. One mobile agent migrates to a designated mobile robot, and drives the robot to the assigned quasi optimal position that is calculated in step 2. Mobile Agents

Host Computer

Assembly Point


Assembly Point MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

70/22 70

SI in Swarm Robotics • Multirobot Control [12] ACO-based Clustering - More objects are clustered in a place where strong pheromone sensed. - When the artificial ant finds a cluster with certain number of objects, it tends to avoid picking up an object from the cluster. This number can be updated later. - If the artificial ant cannot find any cluster with certain strength of pheromone, it just continues a random walk.

71/22 71


SI in Swarm Robotics Start

• Multirobot Control [12] ACO-based Clustering Picking up or not unlocked object is probabilistically determined using the formula:

( p + l) * k f ( p) = 1 − 100 p: density of pheromone= number of adjacent objects. Thus, when an object is completely surrounded by other objects, p is the max. value=9 MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

yes

Have an object

no

Pheromone walk Empty space detected?

Random walk object detected?

no

yes

yes

Put object

no

no

Catch object

Satisfied requirement? yes

End 72/22 72

SI in Swarm Robotics • Multirobot Control [12] ACO-based Clustering

f ( p) = 1 −

( p + l) * k 100

k: constant value Authors selected k= 13 in order to prevent any object surrounded by other eight objects be picked up.

f ( p) = 1 −

(8 + 0) *13 = 1 − 1.04 = −0.04 ≈ 0 100

(never pick it up).


73/22 73

SI in Swarm Robotics • Multirobot Control [12] ACO-based Clustering

f ( p) = 1 −

( p + l) * k 100

l: constant value to make an object be locked. l = zero (not locked). If c=# of clusters & o=# of objects

c p 2 3o c p 1 3o

and and

p f 3⇒ l = 6 pf7⇒l =3 p f 9 ⇒ l =1


Any objects meet these conditions are locked. This lock process prevents artificial ants remove objects from growing clusters. 74/22 74


• Multirobot Control [12] ACO-based Clustering In “pheromone walk” state, an artificial ant tends probabilistically move toward a place it senses the strongest pheromone. The probability that the artificial ant takes a certain direction is n/10, where n is the strength of the sensed pheromone of that direction.

yes

Have an object


no

yes

Put object

no


End 75/22 75



• Multirobot Control [12] ACO-based Clustering An artificial ant carrying an object determines whether to put the carrying object or to continue to carry. This decision is made based on the formula:

p*k f ( p) = 100 The more it senses strong pheromone, the more it tends to put the carrying object. The probability f(p)=1 (must put it down).

yes

Have an object


no

yes

Put object

no


End MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

76/22 76


• Multirobot Control [12] ACO-based Clustering

yes

Have an object

Termination Condition: Pheromone walk

The number of resulted clusters is less than ten, and

Empty space detected?

all the clusters have more or equal

no

yes

to three objects.

Put object

no


End 77/22 77


SI in Swarm Robotics • Multirobot Control [12] Results Airport Field (Objects: 400, Ant Agent: 100) Airport field

Ant Colony Clustering

Specified Position Clustering

Cost

Ave

Clust

Cost

Ave

Clust

1

4822

12.05

1

11999

29.99

4

2

3173

7.93

6

12069

30.17

4

3

3648

9.12

4

12299

30.74

4

4

3803

9.51

3

12288

30.72

4

5

4330

10.82

5

1215

30.31

4

Clust represents the number of clusters, and Ave represents the average distance of all the objects moved. MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

78/22 78

SI in Swarm Robotics • Collective Robotic Search (CRS) [13] A group of unmanned mobile robots are searching for a specified target in a high risk environment.


79/22 79

SI in Swarm Robotics • Collective Robotic Search (CRS) [13] Assumptions - A number of robots/particles are randomly dropped into a specified area and flown through the search space with one new position calculated for each particle per iteration; - The coordinates of the target are known and the robots use a fitness function, in this case the Euclidean distance of the individual robots relative to the target, to analyze the status of their current position; - Obstacle’s boundary coordinates are known.


80/22 80

SI in Swarm Robotics • Collective Robotic Search (CRS) [13] PSO-CRS Algorithm Step 1: A population of robots is initialized in the search environment containing a target and an obstacle, with random positions, velocities, personal best positions (pid), and global best position (pgd). MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

81/22 81

SI in Swarm Robotics • Collective Robotic Search (CRS) [13] PSO-CRS Algorithm Step 2: The fitness value, Euclidean distance from the robot to the target, is determined for each robot where Tx and Ty are the targets coordinates and Px and Py are the current coordinates of the individual robot. fitness =

(T x − Px ) 2 + (T y − Py ) 2


82/22 82

SI in Swarm Robotics • Collective Robotic Search (CRS) [13] PSO-CRS Algorithm Step 3: The robot’s fitness is compared with its previous best fitness (pBestid) for every iteration to determine the next possible coordinate position for each robot in the search environment. The next possible velocity and position of each robot are determined according to the following equations:

vid ( k + 1) = ω × vid ( k ) + c1 × rand () × [ pBest id ( k ) − xid ( k )] + c2 × rand () × [ gBest d ( k ) − xid ( k )] xid ( k + 1) = xid ( k ) + vid ( k + 1) MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

83/22 83

SI in Swarm Robotics • Collective Robotic Search (CRS) [13] PSO-CRS Algorithm Step 4: If the next possible position xid(k+1) resides within the obstacle space, the obstacle avoidance mechanism is employed, otherwise the robot moves to this new position.


84/22 84

SI in Swarm Robotics • Collective Robotic Search (CRS) [13] PSO-CRS Algorithm Obstacle avoidance mechanism:

( horiz , vert , next ) = arc ( x _ start , y _ start , x _ center , y _ center , dir , points ) dir: CW – CCW points=4

nearest corner of the obstacle

The new position of the particle is set at the second point in the arc. MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

85/22 85

SI in Swarm Robotics • Collective Robotic Search (CRS) [13] PSO-CRS Algorithm Step 5: The pBestid with the best fitness for the entire swarm is determined and the global best coordinate location, gBestd, is updated with this pBestid. Step 6: Until convergence is reached, repeat steps.


86/22 86

SI in Swarm Robotics • Collective Robotic Search (CRS) [13] Simulation Results Search space: 20 by 20 units. Vmax = 0.5 units. # of robots = 10 # of targets = 1 # of obstacles=1

87/22 87


SI in Swarm Robotics • Collective Robotic Search (CRS) [13] Simulation Results

Robots’ pathways to the target location. Two robots collide into the obstacle. MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

Robots’ pathways to the target location with obstacle avoidance mechanism. 88/22 88

SI in Swarm Robotics • Collective Robotic Search (CRS) [13] Simulation Results Average Number of Iterations taken by the Swarm with standard deviations over 20 trails Obstacle Type Square

Circle

PSO Parameters 1c1=0.5,c2=2, w=0.6

PSO Parameters 2c1=2,c2=2, w=0.6

Average±std

89.43±9.86

105.6±10.68

Maximum

104

128

Minimum

66

85

Average±std

76.75±8.75

106.1±11.61

Maximum

95

128

Minimum

64

78


89/22 89





90/22 90

Concluding Remarks • Swarm intelligence techniques such as Ant Colony Optimization (ACO) and Particle Swarm Optimization (PSO) have shown to be an effective global optimization algorithms in traditional and swam robotics. • ACO and PSO have been successfully applied in solving problems such as motion planning, gait optimization, group behavior, control, learning, odor-source localization, collective robotic search and mobile sensor deployment, to mention a few.


91/22 91

References [1] E. Sahin and A. Winfield, “Special issue on swarm robotics,” Swarm Intelligence, 2: 69–72, Springer Science, 2008. [2] Pasteels et al. From Individual to Collective Behavior in Social Insects. Pages 155-175, 1987. [3] Richard Vaughan, “Massively multi-robot simulation in stage,” Swarm Intelligence 2: 189–208, 2008. [4] T. Guan-Zheng, H. Huan and S. Aaron, "Ant Colony System Algorithm for Real-Time Globally Optimal Path Planning of Mobile Robots", Acta Automatica Sinica, 33(3), p. 279-285, March 2007. [5] Xueping Zhang, Hui Yin, Hongmei Zhang and Zhongshan Fan, “Spatial Clustering with Obstacles Constraints by Hybrid Particle Swarm Optimization with GA Mutation “, LNCS, Springer, ISBN 978-3-540-87731-8, pp.569578, 2008. [6] Mehtap Kose and Adnan Acan, “Knowledge Incorporation into ACO-Based Autonomous Mobile Robot Navigation”, LNCS, Springer, ISBN 978-3-540-23526-2, pp. 41-50, 2004. [7] Cord Niehaus, Thomas R¨ofer, Tim Laue, "Gait Optimization on a Humanoid Robot using Particle Swarm Optimization", The second workshop on Humanoid Soccer Robots, 2007. [8] Gursel Alıcıa, Romuald Jagielskib, Y. Ahmet Sekerciogluc, Bijan Shirinzadehd, “Prediction of geometric errors of robot manipulators with Particle Swarm Optimisation method”, Robotics and Autonomous Systems 54, pp. 956–966, 2006. [9] Takeshi Matsui, Masatoshi Sakawa, Takeshi Uno, Kosuke Kato, Mitsuru Higashimori, and Makoto Kaneko, "Particle Swarm Optimization for Jump Height Maximization of a Serial Link Robot", Journal of Advanced Computational Intelligence and Intelligent Informatics, Vol.11, No.8 pp. 956-963, 2007. [10] Fabien Moutarde, “A robot behavior-learning experiment using Particle Swarm Optimization for training a neural-based Animat”, 10th International Conference on Control, Automation, Robotics and Vision (ICARCV 2008), 2008. MUSES_SECRET: ORF-RE - © PAMI Research Group of – University ECE493T8:2008-2009 - © Project PAMI Research Group – University Waterloo of Waterloo

92/22 92

References [11] Maki K. HABIB, Hajime ASAMA, “Efficient method to generate collision free paths for autonomous mobile robot based on new free space structuring approach,” In Proceedings of International Workshop on Intelligent Robots and Systems. Japan, November, 1991, 563-567. [12] Yasushi Kambayashi, Yasuhiro Tsujimura, Hidemi Yamachi, Munehiro Takimoto and Hisashi Yamamoto, “Design of a Multi-Robot System Using Mobile Agents with Ant Colony Clustering”, Proceedings of the 42nd Hawaii International Conference on System Sciences - 2009. [13] Smith, L., Venayagamoorthy, G. K. and Holloway, P. (2006): Obstacle Avoidance in Collective Robotic Search Using Particle Swarm Optimization. IEEE Swarm Intelligence Symposium, Indianapolis, Indiana, USA, May 2006. [14] Fei Li, Qing-Hao Meng, Shuang Bai, Ji-Gong Li and Dorin Popescu, "Probability-PSO Algorithm for Multirobot Based Odor Source Localization in Ventilated Indoor Environmen", LNCS, Springer, 978-3-540-88512-2, pp. 1206-1215, 2008. [15] Wisnu Jatmiko, Petrus Mursanto, Benyamin Kusumoputro, Kosuke Sekiyama and Toshio Fukuda, "Modified PSO algorithm based on flow of wind for odor source localization problems in dynamic environments", WSEAS TRANSACTIONS on SYSTEMS archive, 7(2), pp. 106-113, 2008. [16] Wu Xiaoling, Shu Lei, Yang Jie, Xu Hui, Jinsung Cho and Sungyoung Lee,“Swarm Based Sensor Deployment Optimization in Ad Hoc Sensor Networks,“ LNCS, Springer, ISBN 978-3-540-30881-2, pp. 533-541, 2005. [17] Jim Pugh and Alcherio Martinoli, “Multi-robot learning with particle swarm optimization”, International Conference on Autonomous Agents,ISBN:1-59593-303-4, p. 441-448, Japan, 2006. [18] Sabine Hauert, Laurent Winkler, Jean-Christophe Zufferey, and Dario Floreano, “Ant-based swarming with positionless micro air vehicles for communication relay”, Swarm Intell (2008) 2: 167–188. 2008. [19] Yan Meng, Olorundamilola Kazeem and Juan C. Muller, “A Hybrid ACO/PSO Control Algorithm for Distributed Swarm Robots,” Proceedings of the 2007 IEEE Swarm Intelligence Symposium (SIS 2007).


93/22 93

Swarm Intelligence in Robotics

Swarm Intelligence in Robotics

Suggest Documents

Research Advance in Swarm Robotics - Computational Intelligence ...

Research Advance in Swarm Robotics - Computational Intelligence ...

Swarm Robotics - Bio-Inspired Artificial Intelligence - EPFL

Swarm Intelligence

Swarm Intelligence

Evolutionary Swarm Robotics

Swarm Robotics - Springer Link

Swarm Robotics - CiteSeerX

Swarm Robotics - CiteSeerX

Robotica Swarm robotics reviewed

Swarm Robotics Workshop

Juniper Publishers| Robotics Engineering- Swarm Intelligence Algorithms as the Promising

Challenging Applications in Swarm Robotics - Infoscience - EPFL

Task Allocation in Swarm Robotics pdfsubject

Scenarios of Swarm Robotics

An Introduction to Swarm Robotics

Model Independence in Swarm Robotics S. Kazadi

Swarm intelligence

Swarm Intelligence - CiteSeerX

Social Intelligence and Robotics

Swarm robotics: a review from the swarm engineering perspective

Swarm intelligence in intrusion detection: A survey

Swarm intelligence in anomaly detection systems: an

New inspirations in swarm intelligence - Semantic Scholar