See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/226434710

Development of an intelligent agent-based AGV controller for a flexible

manufacturing system

Article in The International Journal of Advanced Manufacturing Technology · March 2007

DOI: 10.1007/s00170-006-0892-9

CITATIONS READS

83 1,885

4 authors, including:

Sharad chandra Srivastava Alok Choudhary

National Institute of Foundry & Forge Technology Loughborough University
39 PUBLICATIONS 2,172 CITATIONS 70 PUBLICATIONS 3,565 CITATIONS

SEE PROFILE SEE PROFILE

Manoj Kumar Tiwari

National Institute of Industrial Engineering
445 PUBLICATIONS 21,084 CITATIONS

SEE PROFILE

All content following this page was uploaded by Manoj Kumar Tiwari on 27 May 2014.

The user has requested enhancement of the downloaded file.

Int J Adv Manuf Technol (2008) 36:780–797
DOI 10.1007/s00170-006-0892-9

ORIGINAL ARTICLE

Development of an intelligent agent-based AGV controller

for a flexible manufacturing system
Sharad Chandra Srivastava & Alok Kumar Choudhary &
Surendra Kumar & M. K. Tiwari

Received: 30 June 2005 / Accepted: 17 November 2006 / Published online: 30 January 2007
# Springer-Verlag London Limited 2007

Abstract Automated guided vehicles (AGVs) are the most the proposed agent-based controller is capable of generating
flexible means to transport materials among workstations of optimal, collision- and deadlock-free path with less com-
a flexible manufacturing system. Complex issues associated putational efforts.
with the design of AGV control of these systems are
conflict-free shortest path, minimum time motion planning Keywords Automated guided vehicles (AGVs) .
and deadlock avoidance. This research presents an intelli- Flexible manufacturing system AGV control .
gent agent-based framework to overcome the inefficacies Multi-agent system (MAS) . Deadlock avoidance
associated with the aforementioned issues. Proposed ap-
proach describes the operational control of AGVs by
integrating different activities such as path generation, 1 Introduction
journey time enumeration, collision and deadlock identifi-
cation, waiting node location and its time estimation, and Flexible manufacturing systems (FMS) aim to combine the
decision making on the selection of the conflict-free productivity of flow lines with the flexibility of job shops,
shortest feasible path. It represents efficient algorithms to attain very versatile manufacturing units achieving high
and rules associated with each agent for finding the operational efficiencies. They are particularly designed for
conflict-free minimum time motion planning of AGVs, low volume, high variety manufacturing, and good decision
which are needed to navigate unidirectional and bidirec- making and management are crucial to maximize the
tional flow path network. A collaborative architecture of benefits that they offer. Stecke and Solberg [1] mentioned
AGV agent and its different modules are also presented. four decision stages for FMS, i.e., design, planning,
Three complex experimental scenarios are simulated to test scheduling and control. An FMS consists of a set of cells:
the robustness of the proposed approach. It is shown that material handling system (automated guided vehicles) and
service centers, etc. Automated guided vehicle systems
S. C. Srivastava : S. Kumar (AGVs) are advanced material handling devices extensively
Department of Production Engineering, used in automated manufacturing systems (AMS) to
Birla Institute of Technology,
transport materials among workstations [2]. The design
Mesra,
Ranchi 835 215, India and implementation of such AGVs require answers to a
number of problems, such as guide path design, controller
A. K. Choudhary devices, and routing algorithms. In the recent past, with the
Wolfson School of Mechanical and Manufacturing Engineering,
emergence of distributive technologies [3] and advanced
Loughborough University,
Loughborough, UK manufacturing paradigms, the AGV design and control
problems attracted the attention of various researchers,
M. K. Tiwari (*) including Tanachoco and Sinrich [4], Egbelu and Tanachoco
Department of Forge Technology, National Institute of Foundry
[5], Egbelu [6], etc. AGVs are under computer control and
and Forge Technology,
Ranchi 834003, India they are the most flexible means to link all the locations of
e-mail: [email protected] the shop floor [7–9], their operations must face some
Int J Adv Manuf Technol (2008) 36:780–797 781

traffic control problem, such as collision prevention and Lastly, the paper ends with concluding remarks and future
deadlock avoidance [10, 11], and minimum time motion research directions.
planning. The objective of this research is to provide an
analytical treatment of some of these issues and to offer a
novel perspective to resolve the issues related to collision- 2 Literature review
and deadlock-free vehicle routing with minimum time
motion. Research articles covering in depth knowledge in the broad
The collision and deadlock-free, shortest time, multiple domain of works related to collision- and deadlock-free,
automated guided vehicle system (MAGVs) are the need of shortest time route and application of AI to the control
today’s FMS. There are varieties of AGVs available in the aspect of AGVs, are considered here for discussion.
manufacturing system, serving a predefined flow path Broadbent et al. [12] coined the concept of conflict-free
network, which consists of a set of nodes interconnected shortest time route. A matrix has been generated by
via set lanes (edge or links). Nodes represent sites for applying Dijkstra’s shortest time path algorithm that
workstation and parking areas for vehicle and serve as pick- describes the occupation time of a vehicle at a node.
up and drop-off points for loads. A demand could be Conflicts at a node (junction) or catching-up conflicts can
generated in the form of pick-up loads at any workstation be resolved by slowing down the vehicle, which is yet to be
and drop off a load to other workstations. The main problem scheduled. The procedure can be applied to any type of
to be tackled here is finding the shortest path of the AGVs path, namely unidirectional and bidirectional, but the
from its present location to the destination station via procedure does not guarantee an optimal solution of
intermediate pickup workstation. There are several possible bidirectional models [13]. Introduced the concept of virtual
paths through which the vehicle can travel, and among these tunnel for the bidirectional path consisting of several
shortest, a conflict- and deadlock-free path is to be adopted segment of multi-unidirectional paths. This concept allows
for routing. The vehicle’s chosen path should be such that, it multiple simultaneous crossing at an intersection.
may not affect the others active travel schedule. Chang and Tanachoco [14] presented an algorithm for
The main thrust of this research is to frame collaborative finding conflict-free shortest time route for AGVs in a
agent-based architecture for real time traffic control of network, which is based on Dijkstra’s shortest path method.
AGVs, with a view to avoiding collisions and deadlocks The main drawback of this procedure is its computational
and achieving minimum journey motion times in an FMS. complexity, as the number of vehicles and number of nodes
It generates a conflict-free shortest path for the AGV in an increases, the response time will increase. Narshimhan et al.
effective manner and, therefore, can overcome the ineffi- [15] analyzed the rerouting of AGVs to encounter inter-
cacies that may arise in complex layout of manufacturing ruption via simulation. But, this procedure does not
system. Agent controller, presented in this paper, controls guarantee an optimal and conflict-free path. Oboth et al.
the AGVs using modules. These modules are associated [16] presented a route generation technique that provided
with rule-based system, heuristics and algorithms. Various conflict-free routes for multiple AGVs with varying speeds.
agents used in this architecture are as follows: guide path It is unconfirmed whether this technique can guarantee the
agent, zone controller agent, journey time database agent, optimal conflict-free route. Recently, several researchers
online traffic controller agent, AGV agent, order agent and [17, 18] have attempted to solve the shop-floor control
interface agent. In addition, contract net protocols are used problems using the concept of agent technology.
for fully automated competitive negotiation through the use Wallace [19] has presented novel agent approach, in
of contracts among different agents. which the rule-based intelligent agents act as traffic
The rest of the article is organized as follows: The managers to allow or disallow mobile robots access to
literature relevant to collision- and deadlock-free minimum points (x-y coordinates within a two-dimensional world)
time motion planning with the application of artificial and segments (lines connecting points) in a system. These
intelligence (AI) is reviewed in Sect. 2. Section 3 describes rules and assessments define the interaction between the
various complexity associated with the problem environ- AGVs and guide-path that controls AGVs movement. As
ment. An overview of agent technology and proposed compared to this distributed approach, the proposed
structure of agent-based AGVs are mentioned in Sect. 4. centralized approach to control the AGV movement,
Agent interaction approaches, various rules, heuristics, and although less flexible, is more reliable and easier in
algorithms associated with different agents are illustrated in implementation as well as it avoids many unnecessary
Sect. 5. Section 6 discusses about the implementation agent-to-agent interactions. Lim et al. [20] suggested a
aspects of proposed model. Three experimental scenarios construction algorithm to design a guide-path network for
are illustrated in Sect. 7. Results of simulation runs and AGVs. Their study used the total travel time, including
effectiveness of the proposed approach are discussed next. waiting and interference time, of the vehicle as the decision
782 Int J Adv Manuf Technol (2008) 36:780–797

criteria for determining the direction of the path segment on implies that the vehicle is to move from its present location
the unidirectional guide-path layouts, and Q learning is (initial node) to a node of a pick-up workstation (interme-
utilized to estimate the travel time of vehicles on path diate node), where the jobs for loading are available. From
segments. Fanti [21] has formulated a zone control scheme there, the loaded AGV proceeds to a node representing a
to tackle the traffic control problems, based on the knowledge drop-off workstation (final node) for unloading, where the
of the AGVs’ operative condition. Lee et al. [22] proposed a AGV is free to be routed again.
systematic two staged traffic control scheme to obtain Distributive traffic control has been adopted in many
collision free minimum time motion of AGVs along loop practical situations, where the range sensors maintain a
less path. Kim et al. [23] modeled an intelligent agent that minimum headway between two adjacent vehicles. When
lies in the two extremes i.e., heterarchical and hierarchical the complexity of the guide path network is high, these
frameworks to solve an industrial warehousing problem. distributive controls resulted in decreased efficiency of the
Yamashita [24] proposed two empty vehicle dispatching system. In these situations, it is difficult to find a substitute
policies, and numerically calculated the waiting time path, and deadlock occurs more frequently with an increase
distributions of the items for each policy using a state space in the number of AGVs. In this research, we have
reduction technique for Markov chains. considered centralized traffic control scheme for multiple
In this research, the main focus is to resolve some of the AGV systems with complicated bidirectional guide paths.
complex issues concerning the collision and deadlock On shop floors, the exact position of an AGV can be
avoidance with minimum time motion planning pertaining obtained only when they pass over pre-specified control
to operational control of AGVs. This paper presents an multi- points fixed on the guide path. Hence, the zone blocking
agent-based framework representing a zone controlled AGV technique, which permits only one vehicle in a given path
environment, incorporating various issues like path genera- segment at a time is suitable for multiple automated guided
tion, link occupation time, collision and deadlock avoidance, vehicles under centralized traffic control [25, 26]. In such a
and suggests the waiting node, waiting time, positioning of system, the guide path is divided into a set of many small
the idle AGVs, pick-up and drop-off nodes associated with an path segments; these segments are termed as a link of the
shortest conflict-free path selection. However the intelligent network. The link from node x to one of its adjacent node y
agent-based framework can be used to overcome the short- is denoted as (x, y). For this link, node x is said to be
comings associated with previous approaches and incorporate adjacent to node y and vice versa. For each pair of adjacent
many beneficial facilities explained as follows. nodes, link (x, y) is said to be unique. For bidirectional path
segment two links (x,y) and (y,x) are considered separately,
i. It is a complete framework to develop a minimum time
and (y, x) is said to be the inverse link of (x,y).
motion plan for AGVs and to govern the AGVs as same
In this paper, we have adopted some implicit assump-
plan.
tions to map the shop-floor scenario. Vehicles can start and
ii. The system provides collision- and deadlock-free AGV
stop only at nodes. In order to avoid collision, temporary
movement and can solve the breakdown problems (if any).
stay at some nodes is permitted. A spin turn of AGVs on a
iii. This centralized MAS control is more reliable and easy
guide path is assumed to be avoided. It is also assumed that
to implement onto shop floor. The shop-floor controller
a vehicle path cannot contain any loops or partial paths,
acts as mediator among other shop-floor activities and
whose start node is the same as its goal node. A rapezoidal
AGV movements that can act to establish seamless co-
velocity profile is used, and the maximum speed for each
ordination among each activities of the shop floor.
profile is fixed at its maximum vehicle speed Vmax
iv. Pre-detected time of reaching the product at its target is
multiplied by velocity parameter ηxy (0<ηxy ≤1) that is
quite important for establishing co-ordination among
assigned to each link. Acceleration and deceleration of
shop-floor activities.
AGVs are assumed to be constant, and can be denoted as
v. The system can not only serve the AGVs on complex
αacc and αdec. Disruption and rescheduling of any active
guide-path but also efficiently control the movement of
travel schedule for other vehicle is not permitted.
AGVs even if the number of AGVs is increased.
Journey time for a path P [I, F] can be determined using
the aforementioned variables. Here (x, y)∈P [I, F], for link
(x, y), when AGV move over it, the time require to clear
3 Shop-floor environment (x, y) equals to (ηxy. Vmax)−1. lxy. Here, journey time for a
path P is represented by:
The problem considered here is related to the dynamic X 1 X X X
environment of flexible manufacturing systems (FMS), τ ð pÞΔ ¼ ηxy ν max  l ðx; yÞ þ θi þ φj þ ξr . . .
i j r
where fixed numbers of AGVs are available to handle
material transfer requests. An order or demand for a vehicle ð1Þ
Int J Adv Manuf Technol (2008) 36:780–797 783

where rationality. Furthermore, agents may have high and low

level reasoning capabilities [29]. In this paper, we have
defined an agent as a computational entity that acts on
θi Additional time for ith acceleration.
behalf of others is autonomous, both proactive and reactive,
φj Additional time for jth deceleration.
and exhibit certain degree of ability to learn; cooperate and
ξr Additional time for rth temporary stay.
can move after receiving the details of certain task. An
P [I, F] Path P from I to F, where I is the initial or start
agent acts autonomously following certain algorithm.
node and F is the final or goal node
Based on their skills, an agent tries to attain the goal
τ(P) Journey time for a path P
defined by the assigned task in a proactive manner. In order
Vmax Maximum velocity of the AGV
to achieve the goal, agent must be mobile, collaborative,
ηxy Velocity parameter associated with the link
and communicative to share their knowledge.
(0<ηxy ≤1)
lxy Length of link (x,y).
4.3 Agent-based architecture of AGVs
Entry time and exit time, associated with each link, is
known using journey time. Suppose an AGV is moving Three basic elements of agents have been identified:
along a path P [I, F] then entry time of (x, y) is estimated network interface, local knowledge model and domain
using τ[{I. . . x}] while the exit time is calculated as knowledge model. Network interface consists of the agents
τ[{I,..P, y}]. A very small response time is also considered. to the network. Various resources like Java classes, other
This is the time from the start of the algorithm execution to agent’s address, message language KQML, ontology and
the start of the vehicle movement. It includes the strategies comes into the purview of local knowledge
computational time of the algorithm and the communication model. Domain knowledge model comprises of dexterity
delay needed to transmit the travel schedule to the vehicle. essential for an agent to perform functional tasks and skills
that may be the method for activating action corresponding
to the received request. Figure 1 represents schematic
4 Background information on agent technology structure of an agent.
In order to implement the agent characteristics in to
4.1 Definition of agent AGVs following functional modules are required: commu-
nication interface, perception, social knowledge, self
It is well documented that there are sundry definitions of knowledge, domain knowledge, learning, reasoning, prob-
the term agent reported by several researchers in the recent lem solving interface, coordination, planning scheduling
past. Fisher [27] defined agent as an encapsulated entity and control, conflict management, application interface, etc.
with embedded AI capabilities. Jennings and Wooldridge In this paper, we have developed a collaborative architec-
[28] defined agent to be an autonomous problem-solving ture incorporating the features from other architectures and
entity, which by nature is communicative, reactive, and goal thus provide powerful and promising capabilities. The
oriented. Davidson et al. [29] mentioned that agent is an internal architecture for AGV agent is shown in Fig. 2.
entity with an ability to interact independently with its
environment through its own sensors and effectors [30]. 4.4 Communication protocols used in agent network
Nwana and Ndumu [31] defined an agent to be a software
and/or hardware component capable of acting rationally, in The protocol aims to efficiently and effectively distribute
order to accomplish tasks on account of its user. Huang and the task among the different agents to ensure the comple-
Nof [32] suggest that an agent is a collaborative computing tion of task smoothly and efficiently. It allows agent to
system capable of reacting autonomously and responding share the information therefore determine the overall
reflexly to the impacts from the environment in a goal behavior and organization of MAS (multi-agent system).
oriented paradigm. Moreover, literature available on role Figure 3 shows the mechanism of sending and receiving
and definition of agents are vast and beyond the scope of information tasks through communication bus (internet/
this section. More details about agents can be found in intranet).
[33–37]. In this proposed MAS, all agents communicate over the
Internet/intranet via a bundle of knowledge query and
4.2 Agent properties and types manipulation language (KQML) messages to transfer data
among each other. Agent communication language (ACL)
According to Jennings and Wooldridge [28], an agent can acts as formalism for exchanging messages. It consists of
exhibit autonomy, social ability, responsiveness and proac- three layers: a content layer, a message layer and a
tiveness, in addition to adaptability, mobility, veracity and communication layer. The actual message is specified by
784 Int J Adv Manuf Technol (2008) 36:780–797

INTERACTION WITH OTHER AGENT AND ENVIRONMENT

NETWORK INTERFACE

COMMUNICATION INTERFACE

MESSAGE QUEUE OF TASK RESPONSE

DEVICE

DOMAIN KNOWLEDGE LOCAL KNOWLEDGE

DEXTERITY ONTOLOGY
SKILLS LANGUAGE
STRATEGY

DATA BASE

Fig. 1 Agent’s architecture

Contract Net Protocol

Communication interface

Coordination Control Conflict Mgmt

Planning/Scheduling Problem Solving Models

Knowledge Domain

Fig. 2 Architecture for AGV agent

Int J Adv Manuf Technol (2008) 36:780–797 785

the content layer. The message layer comprises performa- Based on these factors link occupation time is estimated
tives which are provided by the language (e.g., tell, reply, by JTD agent.
decision making and ask-if), which correspond to the 2. 4.6.2 Journey time database (JTD) agent 1. The journey
speech act from the speech act theory [36]. The performa- time database agent enumerates the link occupation
tive header defines what the message means and what the time of each AGVs. Link occupation time is the
recipient agent should do. For the implementation aspect of interval between the entry time and exit time of a link.
the proposed structure, small subsets of performatives have It also includes the response time that is nothing but the
been customized, as described in Table 4. When a message time from the start of the algorithm execution until start
is transmitted among the agents, it is wrapped by the of the vehicle movement. In Sect. 5, an illustration has
KQML in a standard format. The destination agent can been cited for estimating the link occupation time.
compose the KQML and retrieve the embedded message. These informations are utilized by zone controller agent
The agent language (AL) is developed for each agent to to plan the trajectory along a candidate path.
understand the message it receive and execute the task as 3. 4.6.3 Zone controller (ZC) agent 1. A zone controller
message ‘tells’. agent utilizes link occupation table data to determine a
collision- and deadlock-free trajectory of a vehicle.
4.5 Agent’s functionality Hence, a ZC agent is mainly responsible for trajectory
planning of AGVs. However, it is considered that
This paper presents an idea consisting of society agents to potential collision occur if its link occupation time
resolve the issue related to conflict, deadlock and interrup- overlaps any active occupation schedule by other
tion occurring in a guide path network. Six types of agent vehicle. In this case, ZC agent searches for temporary
have been proposed to develop the architecture of the staying node and time. In order to avoid collision/
control mechanism of AGVs. Each agent is associated with deadlock, ZC agent is responsible for making two types
modules and these modules follow some rule bases, of decisions as discussed in Sect. 5.
heuristics and algorithms. Figure 4 represents the architec- 4. 4.6.4 Online traffic controller (OTC) agent 1. OTC
ture of a proposed multi-agent system framework and agent determines the overall motion planning of
coordination, and negotiation among these agents has been AGVs. This agent is also termed as decision maker.
demonstrated in Table 4. The OTC agent on the basis of communication takes
These agents are described as follows: decision related to the shortest feasible path with
other agent and heuristics and rule bases associated
1. 4.6.1 Guide path (GP) agent 1. The guide path agent is
with them. Details of rules and heuristics are
responsible for finding the possible paths through
described in Sect. 5. After deciding the shortest
which the AGVs can be routed. A guide path agent
feasible path, OTC agent instructs the AGV agents to
receives information from an AGV agent regarding its
initiate its motion and continuously governs its move-
present location, location of pick-up, and drop- off
ment. If any problems related to the AGVs control
workstation. GP agent generates shortest feasible path
(probable location to be head-on collision of AGVs,
and K shortest feasible path based on high level
breakdown of AGVs) arise, it reports the shop-floor
reasoning capability and different modules associated
controller to heal up the trouble.
with it. The working methodology and generation of
5. 4.6.5 Order agent (OA) 1. As generation of new order
shortest path and K shortest path is illustrated in Sect. 5.
to transfer the supplies from one station to another

Fig. 3 Agents sending and re-

ceiving information/tasks Communication Bus (Intranet)
through a communication bus
Task Task Task Task
arrives departure arrives departure

Seek collaborator through TVP

Agent 1 Agent 2

Error information through TVP

Task Task
Task Task arrives departure
arrives departure

Communication Bus (Internet)

786 Int J Adv Manuf Technol (2008) 36:780–797

Fig. 4 Agent-based system ar-

chitecture using proposed ap- JTD Agent
proach, (dashed line shows the
Internet-based communication,
and solid lines show the Intranet
based communication) OTC Agent Interface Agent ZC Agent

GP Agent

Order Agents

AGV Agent 1 Station 1

AGV Agent 2 Shop floor controller Station 2
: ( scheduler …… controller) :
: :

station or entry of new supplies in the system arise; AGV agent, order agent, JTD agent, ZC agent, and OTC
shop-floor controller detects the requirement of AGVs agent.
to transport the supplies from station to station. It The interface agent is responsible for the management of
instructs the order agent to develop a plan for transpor- interaction and the conflict resolution in the agent commu-
tation of the supplies. nity (GP agent, order agent, JTD agent, ZC agent, and OTC
6. 4.6.6 AGV agent 1. Each AGV in the system is referred agent). It contains the knowledge about capability and
to as an AGV agent, and its routing plan is managed by function of each connected agent. Therefore, the message
OTCA. It negotiates with an OTC agent at every can properly pass through the interface agent The interface
incident occurred with AGVs (loading, unloading, agent routes the request/response received to the appropri-
information to cross of each node, breakdown, etc.). ate agent based on its knowledge about the capability of
A detailed structure of AGV agent is outlined in Sect. 4. each agent.
Figure 2 represents the architecture of AGV agent.

4.6 Agent-based system architecture 5 Agent interaction approaches

Once the agents are available, there are two possible An AGV system is based on the guide path constructed of
architectures from which to choose: (1) when all the agents nodes and links having some rules, heuristics and algorithm
handle their own coordination or (2) a group of agents rely associated with the agents to control the AGVs. It involves
on special system programs to achieve coordination. The a natural choice that agent should follow to select the rules
disadvantage of this first architecture is related to the and heuristics inherent in them. This section presents
communication overhead, especially the scalability require- heuristics and rules associated with agents. The working
ment, which is essential for better communication among methodologies of each agent, along with the interaction
agents. As a consequence, the latter federated approach is mechanism among them are also discussed here.
preferred in which agents do not communicate directly with
other agents. Instead, they communicate through a system 5.1 GP agent
program called interface agents. This leads to the develop-
ment of a framework where the agents form a federation in Two algorithms are associated with this agent. The
which they surrender their autonomy to their centralized functioning of guide path agent is based on these
agent. The centralized agent takes the responsibility of algorithms. These algorithms are termed as the shortest
fulfilling their needs. feasible path (SFP) algorithm [38] and K- shortest feasible
The proposed system is shown in Fig. 5, which path (KSFP) algorithm [39, 40]. These algorithms find the
consists of seven components: interface agent, GP agent, shortest and K shortest feasible path from a start node I to
Int J Adv Manuf Technol (2008) 36:780–797 787

6
5 4 notations, which are frequently used in the description of
shortest feasible path and K ,shortest feasible Ppath
F algorithm, are listed in Table 1.
These algorithms are discussed as follows.
3

5.1.1 Shortest feasible path (SFP) algorithm

I 1 2  
SFP Algorithm [42] proceeds to grow like a tree Γ NCt ; LtC
0,I
using the original network G (N,L) until the shortest
feasible path from I to F is found. Let a label dist. ðpn Þ
for each pn 2 NCt denote the feasible path length from 0s to
I, 1 pn . A priority queue £ is also defined, in which the leaf
nodes in ΓI are sorted in their order of Dist (.). SFP
1,2 algorithm is delineated as follows:
1,7
[SFP Algorithm]
Initialization Step: let 0I∈Nc be the only node included
7,5 2, 3 in ΓI=>Nc={0I},
Set dist. (0I)=0, Let £={0I}
Step 1 Among the leaf nodes of ΓI, minimum Dist(.)
5,4 3, 4
value is picked up. Choose arbitrarily if tie occurs.
Let the leaf node picked up be pn , now remove pn
from £.
4,5
4,3 pn is chosen in such a manner that, pn 2 £  NCt
and Dist ðpn Þ ¼ minðDist ðun ÞÞ
5,7 un 2 £  
5, 6 pn is excluded from £, hence, pn 2 NCt  £
Step 2 If n=F then stop, shortest feasible path is p½0I ; pn 
3,2 in ΓI.
Step 3 Expand pn ; nm a child node of pn is generated by
6,F selecting m’s from Adj (n). All the m’s are not
2,1
selected even if m is feasible from p via n, but only
Fig. 5 An example for finding the shortest feasible path via SFP those are considered as the base node for the child
algorithm node of pn that met the criteria given below:

goal node F, respectively. Agent-based control of AGVs Table 1 List of notations used in SFP and KSFP algorithm
requires construction of a set of multiple feasible solutions,
which are called path table [22]. Hence, path table PTI/F for Notations Definition
a start node I to goal node F is a collection of candidate G (N, L) : Linkage-constrained network
path that contains the shortest feasible path and 2nd, 3rd .... GI (Nc, Lc) : Induced network of a given linkage constrained
Kth shortest loop less feasible path. Here, K is an input network G (N, L) with start node I
parameter defined by the operator. These algorithms are Adj (n) : Adjacent of node n
based on the principle of partitioning the solution space d (pn, nm) : Distance between pn and nm
[41]. In this paper, a concept of an induced network model Π1 : Shortest feasible path
Πi : ith Shortest feasible path
has been visualized to find the shortest feasible path instead
Sc : Set of candidate path
of a linkage-constrained network. For a detailed study of B[n] : {(n, m)} where m∈Adj(n)
linkage-constrained network and induced network model, it = : Set of feasible path from start node I to goal
is advised to refer [22, 41]. The intension path P for a path node F
p½0I ; pn  ¼ f0I ; Im::: ; pn g in GI (Nc, Lc) is defined as the path φ : Null set
{I,m,...p,n} on G (N, L) constructed from the base node of  I, pn) 
P(0 : It is defined for each (p,n)∈L, a new element in Nc
Γ NCt ; LtC : Growth of a tree using original network G (N,L)
each element in P. The intension path of any path on GI
(Nc, Lc) is a feasible path on G(N, L) and vice versa. 0I : Start node
Dist(pn) : A label denoting feasible path length from 0I, to pn
Hence, graph search that performed only on GI (Nc, Lc) will
£ : Priority queue
deal with only feasible path in G (N, L). Some of the
788 Int J Adv Manuf Technol (2008) 36:780–797

For m2Adj(n), let nm 2 £  NCt if: Shortest path among Sc is chosen. Let it be Πi+1[I, F].
m is feasible from p via n, except the case pn ¼ 0I , Eliminate Πk+1[I, F] from Sc.
In ΓI, m is not a base node of the ancestor node of Step 3 Check if k=K stop; otherwise k=k+1 and go to step 1.
pn , After applying KSFP algorithm, one constructs a
nm is not repeated in ΓI, i.e., nm2= NCt yet. path table. Each table contains K loop less path
For such nm , let ðpn ; nm Þ 2 NC , and Dist ðnm Þ ¼
t
having same start and goal nodes. The K loop less
Dist ðpn Þ þ d ðpn ; nm Þ. path in the table is shortenend and stored in data
Step 4 If £=φ stop. Otherwise return to step 1 as there is files in order of their journey times. Now, all listed
no feasible path from I to F. paths are referred to the journey time database
agent to determine the link occupation time
Here, it is pertinent to mention that the SFP algorithm behaves associated with each link in corresponding path.
as greedy algorithm. Figure 5 illustrates the approach for
finding shortest feasible path via SFP algorithm. Although it
seems strange that from the star node I to goal node F, there 5.2 Journey time database (JDT) agent
is no direct link, yet it is a part of the KSFP algorithm to
eliminate one or more links from the original network. The journey time database agent generates link occupation
time data according to vehicle speed. Link occupation
time is the interval between the entry time and exit time of
5.1.2 KSFP Algorithm a vehicle on a link. Link occupation time for every link is
stored in the form of link occupation table (LOT). Hence,
This algorithm is evolved keeping in mind the principle of after dispatching of a vehicle, the link occupation
partitioning the solution space provided that initial shortest schedule of the vehicle is stored in LOT. ZC agent
feasible solution exists that can be found from SFP algorithm. utilizes LOT to determine collision free trajectory of a
Algorithm KSFP is delineated as follows: vehicle. Figure 6 represents the proposed data structure of
[KSFP Algorithm] each link.
Initialization step: Using the SFP Algorithm, find the Some preference rules are also considered while consid-
first shortest path, set k=1, Sc=φ Find Π1[I, F]. ering the motion of AGVs. These preference rules are
Step 1 [Generation of feasible deviation path: Shortest delineated as follows:
feasible deviation path from y ij ; j ¼ 1:::Π I which – AGV will move through the straight edge rather than
branches at jth node of the kth shortest feasible two curved situations.
path is found out. Put the paths thus found in Sc]. – AGVs will move through the curved edge rather than
Loop 1:For jth node of Πk [I, F] ( j=1,.... πk), two straight edges.
{ – Before departing from initial node, the vehicle will
Define B*[n] as a subset of B[n]. Let (nk( j),m)∈B* [nk, j], acquire the node position.
if nk( j, m) is not repeated in Π1[I, F],...., Πk[I, F] and
[Link Occupation Table (LOT) Algorithm]
furthermore m is feasible from nk( j−1) via nk(j) when j≠0.
Loop 2:For each [nk( j),m]∈B* [nk , j] Step 1 Refer the path table and determine the set of links
{ at which link occupation time is to be calculated.
Set d(ni(i), . )=∞∀I... ..., j−1. Initial link refers to that link whose initial node
Set d(nk( j), .)=∞ except for d(nk( j), m). contains the idle vehicle.
Find Π1(nk,( j), g) by applying SFP algorithm. Step 2 Input the entry time of vehicle at the initial link
Return d(. , .)’s that are set to ∞ in 1) and 2) to the Step 3 Equation 1 is used to determine the exit time, or
former values link occupation time of the vehicle.
Π1(nk,( j), g) does not exist return to 1) for the next (nk, Step 4 Enclose the entry time and exit time in [ ].
( j),m) Step 5 Refer to the path table, determine the next link
Include {I,...nk( j )}+Π1((nk,( j),F) in Sc (if it is not already where the vehicle has to proceed further.
in Sc ), where {I,...nk( j ) } is the sub path of ΠI[I, F]. Step 6 Repeat step 1 to step 5, until the completion of link
} // end of Loop 2 occupation data related to all the links of selected path.
} // end of Loop 1 Step 7 ∞ is placed at exit time of final link and 0 is placed
Step 2 Estimate Πk+1[I, F], find the shortest path in Sc, at link number.
and set the path as Πk+1[I, F], ∞ Shows the unknown idleness of the vehicle at final
Check if S c =φ, stop, since Π k+1 , ... , Π K nodes loop.
distinguished from Π1, . . ., Πk do not exist. 0 Shows unassigned link for the movement of AGVs.
Int J Adv Manuf Technol (2008) 36:780–797 789

Start Node x
5.3.2 Intersection node rule

Goal Node y These rules are delineated as follows:

Length of Link L (x, y) – Neighbor links of physically occupied link have also to
be considered as occupied. It is necessary to avoid
collision at intersection nodes and also guarantee
Entry time I1 (x,y) safety.
– Entry of the two vehicles can be made possible at any
Space for the 1st occupation Exit time O1 (x,y) node if both the vehicle have to travel on different edges.
AGV occupying Index – Entry of the vehicle at the same or different time is
possible at any node but the link occupation time
: constraint must be observed in case the vehicles have
: to travel on the same link.
:
: A potential collision occurs if the link occupation times
overlap by any occupation schedule of other vehicle. In this
Entry time I N (x,y) case, ZC agent estimates the temporary staying node and
node occupation time. Based on the type of collision,
Exit time O N (x,y) temporary staying node and time are determined. In this
Space for the Nth occupation
AGV occupying index work, we have considered four types of collision. Figure 8
illustrates the schematic representation of these four types.
Each type of collision determines different temporary
I j(x,y)= the jth entry time for link (x,y)
O j(x,y)= the jth exit time for link (x,y)
staying nodes and times. In this case, hitting of a vehicle
Fig. 6 Data structure for link occupation table from rear by a faster moving vehicle is not allowed as η
determines the speed of the vehicle on link.
An illustrative example has been expressed for the Figure 9 represents an illustration for trajectory planning
calculation of link occupation time and is represented in of head on collision case. AGV 1 is dispatched along the
Fig. 7. If an AGV travel from node x to node y, its adjacent path Π1=Π1{1,10}={1,2,... 10}. AGV 2 is to be scheduled
node y during a time interval (tx, ty), the link (x,y) , and its and its candidate path is Π2=Π2{11,15}={11, 12, 8, . . . 4,
inverse link (y,x), and all the links and inverse links starting 14, 15}. Suppose that occupation time of AGV 2 overlaps
at x or y are considered to be occupied during (tx, ty), such a with occupation time of previously dispatched AGV 1.
mechanism would protect vehicles from collision at the Then, temporary staying node and time of AGV 2 is
intersection node. determined using trajectory planning algorithm as follows:

5.3.3 Trajectory planning

5.3 Zone controller agent (ZCA)
[TP Algorithm]
The zone controller agent is responsible for determining the
conflict and interruption free path/route associated with other Step 1 Refer to LOT find the possible collision on any
moving vehicle with in the horizon of journey time schedule. link in ∏2. Trajectory planning is completed if no
The ZC agent determines the collision free trajectory along the potential collision is detected. For the selected case
K candidate path in the table that contain K-shortest feasible detected link is (6, 5).
path from current node to goal node. In order to avoid Step 2 Get the index of AGV occupying that link. For this
collision, LOT is exploited by ZCA to ensure that two case, AGV 1 occupies (6, 5).
vehicles do not occupy a link. Relevant rules associated with Step 3 Take inverse sequence of ∏2 (inv∏2), for this case
ZC agent are as follows. inv∏2=(15,14,4 . . . 8, 12,11)
Step 4 Choose start node (β) of the link that bears
5.3.1 LOT rule potential collision (β=6 for (6, 5)). The tempo-
rary staying node is determined as the first
Overlapping of link occupation time is avoided; ZCA element ξ∈inv∏2 after (β) such that ξ∉∏1 in this
checks the link occupation table and plan in such a manner case ξ=12.
that link occupation times of the vehicle do not overlap Step 5 Pick (ξ) from ∏2, Denote expected entry time by
with other vehicles. later dispatched AGV, for this case (12,8) is
790 Int J Adv Manuf Technol (2008) 36:780–797

Fig. 7 An example representing F Goal Node

calculation of link occupation
time

2 3 4 5
1
Start Node (I)

Passing time at each node

Node 1 = 0: 00: 00 Node 2 = 0: 00: 10 Node 3 = 0: 00 : 18

Node 4 = 0: 00: 26 Node 5= 0:00: 36 Node 6 = 0: 00: 50
Node g = 0: 00: 58

Link (x,y) Entry time of Ii(x,y) Exit time of O-i(x,y)

(1,2), (2,1) 0: 00: 00 0: 00: 18
(2,3), (3,2) 0: 00: 00 0: 00: 26
(3,4), (4,3) 0: 00: 10 0: 00: 26
(2,7), (7, 2) 0: 00: 00 0: 00: 18
(2, 8), (8, 2) 0: 00: 00 0: 00: 18
(3, 7), (7, 3) 0: 00: 10 0: 00: 26
(3, 8), (8,3) 0: 00: 10 0: 00: 26
(4, 7), (7, 4) 0: 00: 18 0: 00: 36
(4, 8), (8, 4) 0: 00: 18 0: 00: 36
(4, 5), (5, 4) 0: 00: 18 0: 00: 50
(5, 6), (6, 5) 0: 00: 26 0: 00: 58
(5, 9), (9, 5) 0: 00: 26 0: 00: 50
(6, 9), (9, 6) 0: 00: 36 0: 00: 58
(6, g), (g, 6) 0: 00: 36 ∞

chosen and I*(12,8). Among the exit times Oi [Overall motion planning algorithm]
(12,8) (i=1,2 . . . N) for link (12,8), select i in
Step 1 initialize i=1
such a manner that Oi (12,8) is the smallest exit
time that is greater than I*(12,8). Step 2 Communicating with ZC agent, plan the trajectory
Step 6 Difference in the expected entry time and along the path ∏i
smaller exit time gives the temporary staying Step 3 Check for ∏i, if there is no potential collision
time (ξr=Oi (12, 8)−I*(12, 8)). In order to occupy detected, i.e., node staying time=0, the collision
this temporary stays modify the trajectory of free minimum time motion path lies among ∏k
AGV 2. (1 ≤ k ≤ i). The smallest τ(Π) among ∏1, ∏2,
Step 7 Calculating φi, ϕj for the temporary stay. ∏3...∏i is selected.
Step 8 Detect if there is any potential collision, then Step 4 If node staying time is greater than zero, and i<I
repeat step 1–7, else stop. for ∏I, go to step 2 and increment i=+1.
Step 5 If node staying time is greater than zero and i=I;
the path ∏k (1≤k≤i) that has the smallest ∏k
5.4 Online traffic controller agent
among ∏1, ∏2, ∏3,...∏i is selected. Step 2
determines the trajectory planning.
This agent is also known as decision-maker and
responsible for overall motion planning of the AGVs.
Based on overall motion planning algorithm and selection TIE breaking rule If the tie occurs, that is if the two paths
rule, OTC agent decides the optimal path to be traversed show same reachable time at destination, OTC agent prefers
by AGV. the route having less traffic by consulting ZC agent.
Int J Adv Manuf Technol (2008) 36:780–797 791

AGV 1

Intersection

AGV 2
AGV 1 AGV 2

a b

Goal Node of AGV 1 AGV 1

Staying AGV

Staying AGV

Goal Node of AGV 1

AGV 1

c d
Fig. 8 Representation of four different types of collisions (a) Head on collision, (b,c,d) Intersection collision

5.5 Order agent situation during fulfilling an order. Hence, it is not safe to
execute the order. Order agents are used to prevent such
When order agents receive information about requirement situations. Concept of virtual order comes into picture when
to transfer the supplies from shop-floor controller, OA order is not safe to carry out. In order to execute real order,
passes the information to other agent of the system for AGV is given virtual order to take it to a safe point or an
finding and solving the transportation demand. According alternative load and unload point to start the execution of
to the OTCA instructions, AGVs can load and dispatch the the real order. Order agent works on the principle of accept
supplies. It is not possible all the time that AGV agent an order rule [19] and is delineated as follows.
executes the order, in some cases; it may lead to deadlock

Fig. 9 Determination of
temporary staying node and time AGV 1 1 10
using trajectory planning
algorithm
2
Π (1, 10_)
1

3 9

4 5 6 7 8 AGV 2

13

14

Π 2
(11, 15)

15
792 Int J Adv Manuf Technol (2008) 36:780–797

ACCEPT an ORDER RULE The prime concern of an order some are now commercially available. Java is employed to
agent is to prevent the deadlock situation of an AGV. Order develop the architecture of present agent-based system.
agent checks every order to see whether the order lays down Each agent developed, in this research is based on the
an AGV to its safe point or its goal node without putting AGV underlying framework of Java Agent Template Lite
into deadlock situation. If the order would result in deadlock (JATLite), which is a prototype agent environment devel-
situation then a virtual order is given to an AGV until it is safe oped by Stanford’s agent-based environment group at
to carry out the original order, or the AGV would have to http://www.java.stanford.edu/. It is a set of lightweight
remain where it is until it could fulfill a virtual or a real order. Java packages that can be used to build multi-agent system.
Negotiations among agents play a vital role during the JATLite facilitates the construction of agent, which can
fulfillment of an order. If the blocking situation arises then send and receives the message using the emerging Stanford
AGVs negotiate with each other for the movement of the communication language KQML [43]. It provides the basic
AGV. If this fails then the blocked AGV will try to find information in which an agent registers with an agent
another route or dummy destination to go to so that it doesn’t message router using a name and password to connect / dis-
put itself or other AGVs in a possibility of deadlock. connect from the internet, send and receives message,
Negotiation is based on high level contract net protocol. transfer files and invoke other programs or actions on the
various computers where they are running.
The wide use of agent technology in industry depends
5.6 AGV agent upon the availability of development tools and a platform
that protect developers from the need to develop basic
Each AGV is associated with an AGV agent. AGV agent functionality of each system. Such tools and platform, in
manages AGV movement. These AGV agents manage an turn, presume the existence of standards that reflect the
AGV by initiating enquiries with other agents and by agreement of developers on what basis the functionality
negotiation with other AGV agent. The detailed collabo- should be presented. Some efforts have been devoted to
rative structure of proposed AGV agent is discussed in provide standards for agent-based systems, but no accepted
Sect. 4. An AGV agent makes a decision on the basis of standard can be found for developing agent-based manu-
message sent by OTC agent. AGV agents communicate facturing systems. KQML is intended as a common
with OTC agent at each incident such as AGVs cross the communication language for agent with KIF [44] as a
node, receives supplies and unloading the supplies etc. If common content format. KQML as ACL is used here.
any AGV becomes breakdown or AGVs come into Some traditional standards have also been used in agent-
location to be head-on collision, OTC agent indicates to based system development, such as the common object
the shop-floor controller to recover the AGV or/and to request broker architecture (COBRA) for inter-agent com-
reschedule the movement plan remained journey of munication and STEP [45] for providing the semantics of
supplies. Some of the rules associated with AGV agent messages in manufacturing application. Currently, two
are delineated as follows: consortia have focused on formalizing standards specifical-
ly to support agents. These are the Foundation for
– Waiting rule: After arriving at its destination, AGV
Intelligent Physical Agents (FIPA) and the National
performs the operation of loading and unloading. In
Industrial Information Infrastructure Protocols (NIIIP).
order to complete its operation, it has to be allowed to
FIPA establishes in 1996 as a worldwide consortium,
wait on its current point. A virtual order is given to an
promotes the development of a specification of generic
AGV when it is unable to accept an order. If, it cannot
agent technologies that maximizes the interoperability with
accept the virtual order then it is unable to move
in and across an agent-based application. FIPA has already
anywhere.
produced version −1 of its set specification called FIPA 98.
– Follow other rules: In order to avoid deadlock, AGVs
NIIIP is a consortium of US companies formed to develop
are allowed to follow OTC agent and hence the
open industry software protocols that will make it possible
situation results in greater efficiency of the system.
for manufacturers and their supplier to operate effectively
The following section describes the implementation in a collective way so that they were part of the same
aspect of the proposed framework. enterprise.

6 Implementation aspect of the proposed model 7 Experimental scenario

With the development of agent-based technology, recently a Experiments were conducted with the intention of
number of agent development tools have been reported and analyzing the capability of proposed agent-based frame-
Int J Adv Manuf Technol (2008) 36:780–797 793

work. Three experimental scenarios with different com- plant. From the implementation point of view, some
plexity have been presented to show the robustness of assumptions are made which are described as follows:
the proposed approach. For the first case, a test problem
1. All AGVs are in the system and travelling to a load and
has been conceived to represent the layout of the AGV
unload station. There is no idle AGV at a workstation
system, which maps the scenario described by [15]. In
or home station.
this scenario, the layout consists of 16 links, nodes
2. There is always an order for AGVs to collect or drop.
represent a parking place of an idle vehicle from where a
3. Multiple pick-up and delivery are avoided. After
pick up and drop off load can be carried out to a respective
completion of a delivery, a new order has been selected
work station or any other important areas like battery
from predefined orders. There is no waiting time except
storage etc. in order to carry out the entire transportation
the computational response time that the code takes to
assignment, three AGVs are used. The situation was
move from completion of an order to new one.
simulated to find out the feasible, collision free path for
4. Buffer capacity of load stations is assumed to be
third AGV from initial node 10 to goal node 4 via pickup
infinite.
(intermediate node) 9.
5. For illustration maximum velocity of AGV is assumed
The second experiment has been conceived with
as 1 m/sec, and Velocity parameter (η)=1(straight link),
increased complexity to determine the computational
0.5 (curved link).
burden of the proposed framework. The simulation run
was carried out for a model guidepath with increased Initially, the numbers of AGVs were taken as six. The
number of nodes and links compared to that of previous numbers of AGVs were decreased as the experiment
one. The model guide path tested in the simulation is proceeds. The last simulation was carried out for two
delineated in Fig. 10. AGVs. Numbers of AGV were decreased with a view to
The specifications related to the guide path are described serve for fewer orders and reduced interference/collisions.
as follows: The simulation carried out was compared with industrial
material handling simulation package.
– Number of links=186 (unidirectional)
– Number of links=372 (bidirectional)
– Number of vehicle=3
– Number of nodes=174 8 Results and discussions
– Maximum vehicle speed Vmax =1.0 m/s (for both
directions) Simulation run has been carried out for the three selected
– Acceleration αacc=deceleration αdec =1.0 m/s2 experimental scenario. The code was developed in Java.
– Velocity parameter (η)=1(straight link), 0.5 (curved The PC used for this simulation is of 256 RAM and
link) 256 MB of virtual RAM with 1.8 GHz and Windows XP
environment.
The main task here is to find the computational response For the first experimental scenario, optimal result is
time. In this case, motion planning has been decided for achieved. In this case {(10 9),(9 8), (8 7), (7 4)} is referred
AGV 3, after the scheduling of AGV 1 and AGV 2. The to as optimal path. OTC agent chooses this path to
path of AGV 1 was from node 3 to node 7, and AGV 2 is complete the associated task with it. This Path is also
scheduled for station 4 to station 6. The task is to find the referred as conflict- and deadlock-free path. Hence, AGV 3
schedule of AGV 3 from start node to goal node such that will start from initial node 10 pick up job at intermediate
conflict-free minimum time motion can be achieved. In this node 9 and dropping off the job to final node 4. Its total
case, computational response time varies as the number of journey time was 161 seconds. The computational response
potential collision increase. Experimental condition has time was .01 second. Hence a total of 161.01 second was
been simulated for various start and goal node to find out used by AGV 3 to complete the given task.
the worst case scenario of AGV 3. Second experimental scenario was the example of a
For the third experiment, a guide path from real world complex guide path. In this case, computational response
industrial scenario has been modeled. In this scenario, times for various starts, and a goal node is determined for
robustness of the proposed approach is shown via increased AGV 3. Table 2 represents the computation time required
efficiency of the AGV system. Hence, the important for motion planning of AGV 3 for various combination of
performance measure of this AGV system is the number start and goal node. It is evident from the table that for the
of order completed. Deadlock occurrence and flowablity of worst case scenario, computational time of the proposed
system is also of prime interest. The layout is large in approach is .22 second. It reveals the fact that these times
surface area and represents the scenario of a manufacturing are well with in the range of the intervals of vehicles
794 Int J Adv Manuf Technol (2008) 36:780–797

9 1

8 2

7 3

10 6 4

a ath Load or Unload Station

Fig. 10 Model guide path used in experimental scenario 2 to test the computational burden of proposed approach

request in the most practical multiple automated guided purpose, efficiency may be defined as number of deliveries
vehicle system (MAGVs). Hence, this can be applied to real per hour. A real comparison between the control system
world application. used in industry and proposed AGV framework is not
For the third scenario, a comparative study is made on feasible; hence, a simulation package (AutoMod) is used to
the basis of efficiency of the system. For the present simulate the performance of the actual AGV system used in
the industry while comparing it with proposed approach. It
is found that for the given layout, deadlock and collision
Table 2 Computational response time for various start and goal node
occur more frequently using the simulation package.
Start station Goal station Computational response time (Sec)

Station 1 Station 10 0.14

Number of order served per

30
Station 2 Station 10 0.12 25
Station 3 Station 10 0.22 AutoMod
20
Station 4 Station 10 0.18
hour

15
Station 10 Station 1 0.12 Proposed
10
Station 10 Station 2 0.12 Agent based
Station 10 Station 3 0.1 5 Framework
Station 10 Station 4 0.14 0
Station 5 Station 9 0.16 2 3 4 5 6
Number of AGVs
Station 6 Station 1 0.08
Station 5 Station 3 0.11 Fig. 11 Comparative study of number of order served/hour for given
layout
Int J Adv Manuf Technol (2008) 36:780–797 795

Table 3 Deadlock formation rate with regard to the number of AGVs 9 Conclusion and future scope
used in real system

Number Number of order Number of Deadlock The main aim of this research is to resolve some of the
of AGVs served/hour deadlocks formation rate complex issues concerning the collision and deadlock
predicted avoidance with minimum time motion planning pertaining
to operational control of AGVs in an FMS. This paper
2 14 8 0.57
presents an multi-agent-based framework representing zone
3 17 14 0.82
4 19 24 1.26 controlled AGV environment incorporating various issues
5 24 32 1.32 like path generation, link occupation time, collision and
6 26 39 1.5 deadlock avoidance, suggests the waiting node and time
estimation, positioning of the idle AGVs, identification of
pickup and drop off nodes associated with an optimal and
Whereas, simulating with proposed agent-based controller, conflict-free path selection. How intelligent agent-based
deadlock- and collision-free minimum time motion is framework can be used to overcome the shortcoming
achieved. Figure 11 represents the comparative study of associated with current approach have been discussed. Six
the number of deliveries/hour using these two controllers. It types of agents have been proposed; each agent is
is evident from the Fig. 11 that the number of orders served associated with some rule base and algorithms. SFP and
using proposed controller for less number of AGVs (up to KSFP algorithms are used by GP agent to generate K
5) is more. As number of vehicle increases it remains shortest path. JTD agent generates LOT. The ZC agent
nearly same. This is due to the fact that as time passes takes care of collision and deadlock avoidance. OTC agent
deadlocks were removed in practical application. However, acts as a decision-maker and determines the overall motion
working of the proposed approach on such a complex guide planning of the system. Order agent takes care of accepting
path in real world application is considerable success. It is and rejecting an order or to generate virtual order. AGVs
seen from the graph that lower number of AGVs result in are managed by an AGV agent. In order to show the
higher rate of deliveries and conflict-free path, where as robustness of the proposed intelligent agent-based frame-
deadlock occurs frequently in the real system. Table 2 work, three experimental scenarios with increased com-
shows the deadlock occurrence rate in the real system. plexity are considered. Simulation result shows that
Deadlock occurrence rate is the ratio of number of proposed framework provides an optimal path, less com-
deadlocks predicted and the number of orders served. This putational burden and higher efficiency. Hence a proposed
is the systematic measure of conflict resolution. agent-based controller could be a viable alternative to a
Tables 3 and 4 large and/ or complex and difficult to design AGV system.
A rule-based system is advocated to address the continuous

Table 4 Input and output sequence dataflow in MAS

Step Dataflow Message type Performative Operation

1 OA-IA Demand for transportation of Request Information of transportation requirement goes into the
supplies multi-agent system
2 IA-GPA Request for present situation Request Request to AGVs to provide the information of its location
of AGVs and status
3 GPA-AGVA AGVs information Update_info Collect the information about present location of all AGVs
4 IA-JTDA Manipulation of link Manipulate Calculation of link occupation time as shown in Fig. 7.
occupation time of each
AGVs
5 IA-ZCA Trajectory planning of AGVs Trajectory_plan Trajectory planning of all feasible paths
6 IA-OTCA Decision of shortest feasible path Shortest_path Evaluation of shortest and collision- free path
7 OTCA-AGVA AGV traffic control Control Communicate with AGVs at crossing of each and every
node
8 AGVA-OTCA Fault occurrence Agv_fault When a breakdown occurs, AGV agent informs the OTCA
9 OTCA-shop-floor Inform to shop-floor Fault_recovery Report to the shop-floor controller
controller controller for fault recovery
10 Shop-floor controller Transportation information about Shop-floor controller instructs the order agent to reschedule
instructs the order loading stations to target station the transportation plan for supplies (which could not be
agent reach its destination)
796 Int J Adv Manuf Technol (2008) 36:780–797

routing of AGV, while negotiating unexpected interruptions 17. Choi KH, Kim SC, Yook SH (2000) Multi-agent hybrid shop
on certain edges (like mechanical failure of any movable floor control system. Int J Prod Res 38:4193–4203
18. Lu TP, Yih H (2001) An agent based production control
AGV). Manual operator is adopted on these occasions to framework for multiple line collaborative manufacturing. Int J
counter any failure in its progress. Prod Res 39:2155–2176
We propose few future research areas also. It is expected 19. Wallace A (2001) Application of AI to AGV control agent: agent
that the control approach proposed here could also be control of AGVs. Int J Prod Res 39:709–726
20. Lim JK, Lim JM, Yoshimoto K, Kim KH, Takahashi T (2002)
applicable in another context like a transportation network. A construction algorithm for designing guide paths of auto-
A developing agent-oriented framework with automated mated guided vehicle systems. Int J Prod Res 40(15):3981–
rule generation using an evolutionary neuro-fuzzy system 3994
to negotiate the conflict and interruption related to 21. Fanti MP (2002) Event-based controller to avoid deadlock and
collision in zone control AGVS. Int J Prod Res 40(6):1453–
operating control of AGV, increasing the efficiency of the 1478
proposed approach for larger number of AGVs in FMS. 22. Lee JH, Lee BH, Choi MH (1998) A real time traffic control
scheme of multiple AGV systems for collision free minimum time
motion: a routing table approach. IEEE Trans Syst Man Cybern,
Part A, Syst Humans 28(3):347–358
References
23. Kim BI, Graves RJ, Heragu SS, Onge AS (2002) Intelligent agent
modeling of an industrial warehousing problem. IIE Trans
1. Stecke KE, Solberg JJ (1983) Loading and control policies for a 34:601–612
flexible manufacturing system. Int J Prod Res 19(5):481–490 24. Uzam M (2004) The use of Petrinet reduction approachfor an
2. Maxwell WL, Muckstatd JA (1982) Design of automated guided optimal deadlock prevention policy for FMS. Int J Adv Manuf
vehicle systems. IIE Trans 14:114–124 Technol 23(3–4):204–220
3. Jennings NR, Wooldridges M (1995) Applying agent technology. 25. Huang J, Palekar US, Kapoor SG (1997) A labeling algorithm for
Appl Artif Intell 9:357–369 the navigation of automated guided vehicles. Trans ASME J Eng
4. Tanchoco JMA, Sinriech D (1992) Osl-optimal single-loop guide Ind 115:315–321, Aug 1997
paths for AGVS. Int J Prod Res 30:665–681 26. Miller RK (1987) Automated guided vehicles and automated
5. Egbelu PJ, Tanchoco JMA (1986) Potential fore bidirectional manufacturing. Dearborn, MI, Soc Manufact Eng
guide path for an automated guided vehicles based systems. Int J 27. Fisher M (1994) Representing and executing agent based systems.
Prod Res 24:1075–1097 In: Proceedings of the ECAI’94 Workshop on agent theories.
6. Egbelu PJ (1987) The use of non simulation approaches in Architecture and languages. Springer, Berlin Heidelberg New
estimating vehicle requirements in an automated guided vehicles York, pp 307–323
based transport system. Mater Flow 4:17–32 28. Jennings NR, Wooldridge M (1998) Application of intelligent
7. Viswanadham N, Narahari Y, Johnson TL (1990) Deadlock agents. In: Jennings NR, Wooldridge MJ (eds) Agent technology
prevention deadlock avoidance in flexible manufacturing sys- foundation, applications and markets. Springer, Berlin Heidelberg
tem using Petrinet models. IEEE Trans Robot Autom 6 New York, pp 3–28
(6):713–723 29. Davidsson P, Astor E, Ekdah LB (1994) A framework for
8. Kumar RR, Singh AK, Tiwari MK (2004) A fuzzy based autonomous agent based on the concept of anticipatory systems.
algorithm to solve the machine loading problems of an FMS and Proceedings of cybernetics and systems’94, vol. II. World
its neuro fuzzy petri net model. Int J Adv Manuf Technol 23(2– Scientific, Singapore, pp 1427–1434
3):318–341 30. Maes P (1995) Modeling adaptive autonomous agents. In:
9. Rajotia S, Shanker K, Batra JL (1998) A semi-dynamic time Langton CG (ed) Artificial life: an overview. MIT Press,
window constrained routing strategy in an AGV system. Int J Cambridge, MA, pp 135–162
Prod Res 36(1):35–50 31. Nwana HS, Ndumu DT (1997) An introduction to agent
10. Lee CC, Lin JT (1995) Deadlock prediction and avoidance based technology in software agents and soft computing. In: Nwana
on Petri nets for zone-control automated guided vehicle systems. HS, Azami N (eds) Towards enhancing machine intelligence.
Int J Prod Res 33:3249–3265 Springer, Berlin Heidelberg New York, pp 3–26
11. Reveliotis SA (2000) Conflict resolution in AGV systems. IIE 32. Huang C-Y, Nof SY (2000) Formation of autonomous agent
Trans 32:647–659 networks for manufacturing systems. Int J Prod Res 38(3):607–
12. Broadbent AJ, Besant CB, Premi SK, Walker SP (1985) Free 624
ranging AGV systems: promises, problems and pathways, 33. Wooldridge M, Jennings NR (1995) Intelligent agents: theory and
problems and pathways. Proceedings of the 2nd International practice. Knowl Eng Rev 10:115–152
Conference on Automated Material Handling UK, pp 221– 34. Jennings NR, Wooldridge M (2000) Agent-oriented software
237 engineering. Handbook of agent technology. AAAI/MIT Press
13. Tagaboni, F, Tanchoco JMA (1988) A LISP-based controller for 35. Lou P, Zhou Z, Chan YP, Xi W (2004) Study on multi-agent
free ranging automated guided vehicle systems. Int J Prod Res based. Agile Supply Chain 23(3–4):197–204
26:173–188 36. Singh SP, Tiwari MK (2002) Intelligent agent framework to
14. Chang WK, Tanchoco JMA (1991) Conflict-free shortest time determine the optimal conflict-free path for an automated guided
bidirectional AGV routing. Int J Prod Res 29:2377–2391 vehicle system. Int J Prod Res 40(16):4195–4223
15. Narshimhan R, Batta R, Karwan M (1998) Routing automated 37. Mondal S, Tiwari MK (2002) Application of autonomous agent
guided vehicles in the presence of interruption. Int J Prod Res network to support the architecture of a Holonic manufacturing
37:653–681 system. Int J Adv Manuf Technol 20:931– 942
16. Oboth CR, Batta R, Karwan M (1999) Dynamic conflict-free 38. Mondal S, Tiwari MK (2003) Formulation of mobile agents for
routing of automated guided vehicle. Int J Prod Res 37:2003– integration of supply chain using the KLAIM concept. Int J Prod
2030 Res 41(1):97–119
Int J Adv Manuf Technol (2008) 36:780–797 797

39. Pierce AR (1975) Bibliography on algorithms for shortest path, 43. Finin T, Fritzon R, McKay D, McEntire R (1993), KQML
shortest spanning tree, and related circuit routing problems language and protocol for knowledge and information
(1956–1974). Networks 5:129–149 exchange. Technical Report, University of Maryland,
40. Dung DA, Grover WD, MacGregor MH (1994) Comparison of k- Baltimore
shortest paths and maximum flow routing for network facility 44. Genesereth MR, Fikes RE (1992) Knowledge interchange format.
restoration. IEEE J Sel Areas Commun 12(1):88–99 Version 3.0 reference manual. Technical Report Logic-92-1,
41. Katoh N, Ibaraki T, Mine H (1982) An efficient algorithm for K Computer Science Department, Stanford University, Palo Alto,
shortest simple paths. Networks 12:411–427 CA [http://www.cs.umbc.edu/agents/kse/kif]
42. Topkis DM (1988) A K shortest path algorithm for adaptive 45. Bjork B, Wix J (1991) An introduction to STEP. Technical
routing in communications networks. IEEE Trans Commun 36 Report, VTT Technical Research Center of Finland and WixMe-
(7):855–859 Lelland Ltd, UK

View publication stats