A COLOURED TIMED PETRI NET MODEL FOR REAL TIME CONTROL OF AGV

SYSTEMS

MARIAGRAZIA DOTOLI, MARIA PIA FANTI

Automated Guided Vehicle Systems (AGVSs) are material handling devices

representing an efficient and flexible option for products management in Automated
Manufacturing Systems. In AGVSs vehicles follow a guide-path while controlled
by a computer that assigns route, tasks, velocity, etc. Moreover, the design of
AGVSs has to take into account some management problems such as collisions and
deadlocks. This paper presents a novel control strategy to avoid deadlock and
collisions in zone controlled AGVSs. In particular, the control scheme manages the
assignments of new paths to vehicles and their acquisition of the next zone.
Moreover, we propose the use of Coloured Petri Nets (CPNs) to model the
dynamics of AGVSs and implement the control strategy stemming from the
knowledge of the system state. Additionally, extending the CPN model with a time
concept allows investigation of the system performance. Several simulations of an
AGVS with varying fleet size while measuring appropriate performance indices
show the effectiveness of the proposed control strategy compared to an alternative
policy previously presented by one of the authors.

1. Introduction
Automated Guided Vehicle Systems (AGVSs) are an efficient and flexible option for
material handling in Automated Manufacturing Systems (AMSs) (Kim and Tanchoco 1991,
Viswanadham and Narahari 1992). Each vehicle follows a guide-path under the computer
control that assigns routes, tasks, velocity, etc. The programming capability for path selection
and reconfiguration of AGVSs easily accommodates changes in production volume and mix
of AMSs. However, an AGVS controller has to deal with some traffic problems such as
collisions and deadlocks, that are caused by sharing the same guide-path.
This paper makes use of a standard technique for vehicle management of AGVSs, i.e.,
zone control. More precisely, the guide-paths are separated into disjoint zones and deadlock
occurs when a set of AGVs competes for a set of zones detained by vehicles of the same set.
Deadlock has been widely studied for AMSs and the most used approaches to describe
interactions between parts and resources are Petri nets (Banaszak and Krogh 1990, Ezpeleta et
al. 1995, Xing et al. 1996, Hsieh and Chang 1994, Viswanadham et al. 1990, Wu and Zhou
2001, Chu and Xie 1997, Huang et al. 2001), automata (Lawley 2000, Lawley et al. 1998,
Reveliotis and Ferreira 1996, Reveliotis et al. 1997) and digraphs (Cho et al. 1995, Fanti et al.
1997, Fanti et al. 2000, Fanti et al. 2001, Wysk et al. 1991).
To cope with deadlock in zone-controlled AGVSs, Lee and Lin (1995) use high-level Petri
nets and Yeh and Yeh (1998) introduce a deadlock avoidance algorithm based on a digraph
approach. Such algorithms avoid deadlock but do not prevent a peculiar condition, known in
literature as “restricted deadlock” (Banaszak and Krogh 1990, Cho et al. 1995, Fanti et al.
1997, Wysk et al. 1991, Kumaran et al. 1994). When a restricted deadlock occurs, the system
is not in deadlock condition but some vehicles permanently remain in circular wait, partly
because some of them are blocked and partly since the controller prevents them from moving.
In addition to the above approaches, Reveliotis (2000) solves deadlock problems in AGVSs
using a variation of the Banker's algorithm and a dynamic route planning that performs route
selection in real-time. Wu and Zeng (2002) addressed the deadlock and conflict problems in
the AGVS with unidirectional guide-paths: the coloured resource oriented Petri net model
developed by Wu and Zhou (2001) is modified to apply the control problem to AGVSs and a
simple control law is presented. Moreover, Fanti (2002a, 2002b) proposed some deadlock
avoidance policies based on digraph analysis. Such policies exhibit low computational costs if
the layout of the AGVS is simple, but they can be too restrictive when the AGVS presents a
large number of bi-directional paths.
The objective of this paper is twofold. First, it proposes Coloured Timed Petri Nets
(CTPNs) to model the layout of AGVSs and the paths that the travelling vehicles follow
(Jensen 1992). To this aim, simplifying the approach proposed by Fanti (2002b), the AGVS
structure and dynamics are described by a Coloured Petri Net (CPN): tokens are vehicles and
the path that each AGV has to complete provides the token colour. As several authors point
out (Ezpeleta and Colom 1997, Feldmann and Colombo 1998, Nandula and Dutta 2000),
CPNs allow concise modelling of dynamic behaviours in industrial applications and are
graphically oriented languages for design specification and verification of systems. The
proposed model is resource oriented, i.e., places are associated with resources rather than with
vehicles operational conditions. The resulting distinctive mark of the model is that the CPN
structure can be easily updated if the AGVS layout changes. Moreover, extending the CPN
with a time concept allows the investigation of the system performance. The execution of a
CTPN works in a similar way as the event queues found in many programming languages for
discrete event simulation (Jensen 1992). Hence, the choices of resource oriented CPNs and
2
temporisation are important to obtain a model suitable for applications in simulation and
control of AGVSs as well as performance investigation.
Secondly, the paper proposes a control policy to manage traffic in AGVSs. In particular,
the paper focuses on the specification of the real time controller: this is in charge of taking
decisions on path assignments and vehicle moves, so that collisions and deadlocks are
avoided. Starting from a previous analysis of deadlock performed for AMSs, a deadlock
avoidance policy is presented based on the knowledge of the actual marking of the CTPN.
More precisely, deadlock is avoided in real time on the basis of a one step look-ahead
procedure obtained by a digraph analysis of interactions between zones and vehicles proposed
in Fanti (2002a). However, the procedures to avoid restricted deadlocks presented by Fanti
(2002a) suffer either from high computational complexity, when the AGVS layout presents a
large number of cycles, or from low permissiveness. Hence, in order to improve the AGVS
control strategy, the paper introduces a new procedure to avoid restricted deadlocks. The
proposed strategy works in polynomial time and it is based on the results of a simulation run
of the CTPN model.
Finally, the introduced real time control policy is compared with the most permissive
policy introduced in Fanti (2002a) and some simulation studies show that it exhibits better
performance indices. In particular, the CTPN modelling the system is implemented in the
MATLAB-Stateflow environment (The MathWorks Inc. 1997). Based on Petri nets
representation as finite state automata, a novel technique is devised to execute the CTPN
dynamics in the MATLAB-Stateflow software. The resulting simulation model is compact,
simple to implement and keeps the CTPN modularity feature while reproducing its structure.
The paper is organised as follows. Section 2 describes the AGVS under study and the
proposed control architecture, Section 3 defines the CTPN modelling the AGVS layout and
dynamics. Moreover, Section 4 recalls previous results on deadlock avoidance in AGVSs and
defines the relationships between CTPNs and digraphs. In addition, Section 5 defines the
control policy and Section 6 describes the controller synthesis. Finally, Section 7 presents
several simulation results for the case study and Section 8 draws the conclusions.

2. The AGVS description and control structure

We consider an AGVS layout involving unidirectional and bi-directional guide-paths. The

AGVS is divided into several disjoint zones and each zone can represent a workstation that a
vehicle can visit, an intersection of paths or a straight lane (see figure 1). While straight

3
unidirectional lanes can be modelled by several zones, bi-directional paths should be divided
into a minimum number of zones in order to keep the controller simple. Moreover, the AGVS
includes a docking station where idle vehicles are parked. The set of zones of the AGVS is
denoted by Z={zi i=1,…,NZ}, where zi for i=2,…, NZ represents a zone and z1 denotes the
docking station. In addition, V={vj: j=1,…, NV} denotes the set of vehicles available in the
system. Since each zone can accommodate only one vehicle at a time, zones zi for i=2,…, NZ
have unit capacity. On the contrary, the docking station can accommodate all the vehicles of
the system and is modelled by zone z1 with capacity equal to NV.

Example 1. The AGVS shown by figure 1 connects six workstations (denoted in the figure
by z2, z3, z4, z5 and z6) and the docking station z1. The load/transfer stations located close to
the workstations represent five zones (z1, z2, z3, z4, z5 and z6) and four additional zones denote
the intersections of the paths with parts of lanes (z7, z8, z9 and z10). The paths z1-z7, z1-z10, z10-
z9, z7-z8 and z8-z9 are bi-directional and the others are unidirectional.

z4 z4

z3 z5
z3 z5
z8 z9

z2 z7 z10 z6
z1
z2 z6
z1

Figure 1. A zone-control AGVS where arrows denote guide-paths and dashed lines denote
zone boundaries.

4
 Each vehicle starts its travel from a zone zi, it reaches a selected zone zj where it loads a
part, then unloads it in the destination zone zk and concludes the travel to the docking station.
Hence, we describe the route r(v) assigned to the vehicle v∈V by the zone sequence
r(v)=(zi…zj…zk…z1). In the sequel, rr(v) denotes the residual route that v∈V has to visit to
complete its travel starting from a certain system configuration: obviously, it is a subsequence
of r(v).
The control structure proposed to manage the AGVS traffic avoids deadlock and collisions
and is composed of two levels (Fanti 2002a, 2002b): the path scheduler and the real time
controller. The path scheduler determines the routes to assign to each AGV and selects the
configurations in which it is necessary to assign a new routing to a vehicle. Route planning is
one of the main activities of the AGVS vehicle management (Tanghaboni and Tanchoco
1988, Co and Tanchoco 1991, Rajotia et al. 1998a, Fanti 2002a) and can be faced by two
approaches: static and dynamic planning. Static planning (Ashayeri et al. 1985, Blair et al.
1987, Egbelu and Tanchoco 1984) assigns the flow path with the minimum distance to the
vehicle mission. In this case, the possibility of vehicle blocking is high because route
selection is made regardless of traffic congestion. On the other hand, dynamic planning takes
routing decisions at each node of the AGV journey on the basis of the positions and speeds of
vehicles and of the node occupation times (Walker et al. 1995, Kim and Tanchoco 1991,
Oboth et al. 1999). Hence, to manage traffic congestion, dynamic planning can create loops of
vehicles and requires the introduction of buffering areas such as loops, sidings and spur types
of buffers (Egbelu and Tanchoco 1986, Rajotia et al. 1998a). In this paper, following the
Fanti (2002a) approach, we propose a semi-dynamic routing that limits the traffic congestion
typical of static planning and overcomes the mentioned drawbacks of dynamic planning.
Indeed, the proposed semi-dynamic algorithm can assign different routes to the same vehicle
mission depending on the traffic status of the network. On the other hand, this strategy
adheres the route to the AGV until the mission is concluded. In our control strategy, the path
scheduler proposes the paths and the real time controller finally assigns the route to the
vehicle after checking that the mission can be completed without deadlocks and restricted
deadlocks. More precisely, the real-time controller has two functions: i) it validates the path
proposed to a vehicle that is waiting for an assignment in the docking station or is travelling
within the system to go back to the docking station (path validation); ii) it validates the next
zone in the path to prevent immediate deadlocks and collisions, by enabling or inhibiting the
AGV zone acquisition (zone validation). This paper deals with the real time controller
synthesis, that performs decisions in closed loop on the basis of the system state knowledge.
5
3. The coloured timed Petri net model
Since the real time controller decisions are based on the system configuration knowledge,
it is necessary to formally describe and model the AGVS. In this section we propose a method
to build the coloured Petri net modelling the system layout and dynamics. Petri Nets (PNs) are
a graphical and mathematical technique to model systems concurrency and synchronisation.
As Moore and Gupta (1996) point out, PNs are a concise representation of discrete event
systems capturing precedence relations and structural interactions. However, extensions of
PNs such as Coloured Timed Petri Nets (CTPNs, Jensen 1992) associate tokens, places,
transitions and arcs with attributes and enhance the power of PNs in modelling complex
systems. Hence, CTPNs find their wide application in modelling, analysis and control of
manufacturing systems (Jiang et al. 2000, Ezpeleta and Colom 1997, Feldmann and Colombo
1998, Nandula and Dutta 2000) and of AGVSs (Hsieh 1998, Zeng et al. 1991). In particular,
Zeng et al. (1991) model AGVSs by a CTPN describing the system paths that each vehicle
has to accomplish. Such a CTPN model is built at each proposed routing plan and is used to
simulate the AGVS and to detect deadlocks and collisions. On the other hand, Hsieh (1998)
proposes a CTPN capable of systematically modelling an AGVS taking into account the
floor-path layout and the management policy. Starting from these models, the CTPN
presented here incorporates the description of the AGVS layout, the path that each vehicle has
to follow and implements the traffic management. Moreover, the CTPN structure is
automatically generated and can be changed as the AGVS layout is modified. In addition, the
events relevant for the AGVS behaviour can be easily controlled on the basis of priority rules
and of the system state knowledge, i.e., the marking of the CTPN.

3.1 Overview of coloured timed Petri nets

A CTPN is an 8-tuple CTPN=(P, T, Co, Inh, C+, C-, Ω, M0) where P is a set of places, T

is a set of transitions, Co is a colour function defined from P∪T to a set of finite and not
empty sets of colours (Ezpeleta and Colom 1997, Jensen 1992). Co maps each place p∈P to a
set of possible token colours Co(p) and each transition t∈T to a set of possible occurrence
colours Co(t). Inh is a weight function for an inhibitor arc which connects a transition to a
place. The inhibitor arc between a place p∈P and a transition t∈T (i.e., Inh(p,t)=1) implies
that transition t can be enabled if p does not contain any token.

6
 Moreover, C+ and C- are the post-incidence and the pre-incidence |P|×|T| matrices

respectively, where |X| refers to the cardinality of set X. More precisely, C+(p,t) associates to

each set of colours of Co(t) a set of colours of Co(p). C+(p,t) (C-(p,t)) is represented by

means of an arc from t to p (from p to t) labelled with the function C+(p,t) (C-(p,t)). The set
Ω is defined as follows: Ω=∪x∈P∪T{C(x)}.
A marking M is a mapping defined over P so that M(p) is a set of elements of Co(p), also
with repeated elements, i.e., a multi-set (Jensen 1992) corresponding to token colours in the
place p. M0 is the initial marking of the net. Just like in ordinary Petri nets, we can define the

matrix C=C+-C- and call it the flow matrix or incidence matrix.

A transition t∈T is enabled at a marking M with respect to a colour c∈Co(t) iff for each

p∈•t, M(p)≥C-(p,t)(c): this fact is denoted as M[t(c)>. The firing of a transition leads to a new
marking M’, that is denoted by M[t(c)>M’. A sequence M[t1(c1)>M1[t2(c2)>M3…
Mn-1[tn(cn)>M’ is denoted by M[ω>M’, where ω=t1(c1)t2(c2)…tn(cn) is a firing sequence; in
this case we say that M’ is reachable from M. The symbol R(M) denotes the set of reachable
markings from M.
Now, to investigate the performance of the system it is convenient to extend the CPN with
the time concept (Jensen 1992). To this aim we introduce a global clock, i.e., the clock values
τ∈ℜ+ represent the model continuous time, where ℜ+ is the set of non negative real numbers.
Moreover, the temporisation of Petri nets can be achieved by attaching time either to places,
or to transitions (Desrochers and Al-Jaar 1995), or else to the expression functions of arcs
(Jensen 1992). In the third method, a delay associated to the output arc of each transition
describes the earliest model time at which the token can be removed from the place by the
enabling of a transition, i.e., there is a delay after which the token becomes available. Here,
we specify the time delay by means of the function δ: P×T→ℜ+, where δ(p,t) describes the
earliest time delay at which the token can be removed from place p after the firing of
transition t. Moreover, we allow each token to have a time stamp attached to it, in addition to
the token colour. The time stamp is represented by the function s: Co(p)→ℜ+, where s(c)
describes the time elapsed from the arrival of the c-colour token to the actual place and s(c) is
reset to zero as soon as the token of colour c arrives to the place. More precisely, when a
transition t∈T is enabled at marking M with respect to a colour c∈Co(t) and fires, the token
reaches an output place and remains unavailable during the time delay δ(p,t), after which the

7
token becomes available. Since the time stamp s(c) is reset as soon as the token arrives to the
place, the transition enabled by the considered token is said to be ready for execution when
the stamp satisfies condition s(c)≥δ(p,t).

3.2 The CTPN modelling the AGVS behaviour

The CTPN integrates the Petri net structure modelling zones and paths with the model of
vehicles travelling in the system and describes the AGVS behaviour. Moreover, the time
attribute allows attaching various time-based measures to the system model (Jensen 1992).

In our framework, the coloured timed Petri net CTPN=(P, T, Co, Inh, C+, C-, Ω, M0)
structure describes the AGVS guide-paths in a resource oriented framework. More precisely,
a place zi∈P denotes zone zi∈Z and a token in zi represents a vehicle that is in zone zi. The
transition set T models the guide-paths between consecutive zones. Moreover, the set of arcs
(P×T)∪(T×P) is defined as follows: if zi and zm are two consecutive zones in the AGVS, then
transition tim belongs to T and is such that tim∈•zm and tim∈zi•.
To admit just one vehicle in each zone other than the docking station, there is an inhibitor
arc between each place zi∈P with i≠1 and transition tim∈T, i.e., Inh(zi,tim)=1. Since z1 can
accommodate NV tokens, there are no inhibitor arcs between z1 and each ti1∈T.
Having described the skeleton of the CTPN, it is necessary to model the AGVS behaviour
and the travels of vehicles. Hence, each AGV v∈V is modelled by a coloured token and its
token colour is <rr(v)>, where rr(v) is the residual path that the vehicle has to follow to reach
the docking station. The state of the AGVS is represented by the marking of the CTPN, i.e., if
M(zi)=<rr(v)>, then vehicle v is in zone zi and its colour <rr(v)> represents the sequence of
zones that v has to visit starting from the current marking. Consequently, the colour domain
of place zi∈P is:
Co(zi)={<rr> where rr is a possible sequence of zones and zi is the first zone of rr}.
Moreover, Co associates with each transition tim a set of possible occurrence colours:
Co(tim)={<rr> such that rr is a possible sequence of zones and zi and zm are respectively
the first and the second element of rr}.
Here, the CTPN is represented by the incidence matrix C that contains a row for each place
zi∈P and a column for each transition t∈T. More precisely, each element C(zi,t) is a function
that assigns an element of Co(t) with t∈T to an element of Co(zi) with zi∈P. The incidence

8
matrix is computed as C=C+-C-, where the pre- and the post-incidence matrices C- and C+,
respectively, are defined as follows:
1. if there exists an arc from zi to tim then C-(zi,tim)=“ID”, where ID stands for “the function
makes no transformation in the elements”, otherwise C-(zi,tim)=0. According to this
definition, each token leaving a zone zi∈P is not modified;
2. if there exists an arc from tim to zm then C+(zm,tim)=“UP”, where UP is a function that
updates the colour <rr> with the colour <rr’>, otherwise C+(zm,tim)=0. More precisely, rr’
is the residual sequence of zones obtained from rr disregarding the first element zi.
The set Ω is defined by Ω={Co(x): x∈P∪T}. Moreover, considering that at the initial
marking M0 a route r(v) is assigned to each v∈V, M0 is defined as follows: if zi∈P is the first
zone of a route r(v) for some v∈V, then M0(zi)=<r(v)> else M0(zi)=<0>.
In the defined CTPN each token is characterized by its colour and its stamp s(rr(v)), that
equals the time spent in the current place. Now, let us suppose that after the firing of a
transition tqi at time τ a token representing vehicle v∈V with colour rr(v)=<(zi, zm,…, z1)> is
in zone zi (see figure 2). As soon as tqi fires, the time stamp of v is reset to s(rr(v))=0 and the
token will be available after δ(zi,tqi) time units. Here, the time delay δ(zi,tqi) is the time
necessary for each AGV to move from zone zq to zone zi. Now, a transition tim is enabled if
two conditions are simultaneously verified:
C1) M(zm)=<0> if m≠1.
C2) M(zi)≥C-(zi,tim)(<rr(v)>) with M(zi)=<rr(v)> and rr(v)=(zi zm, …z1).
The first condition refers to the inhibitor arcs, the second one represents the enabling
condition of the CTPN at marking M. Moreover, at time τ+δ(zi, tqi), the time stamp of v is
s(rr(v))=δ(zi, tqi) and tim is ready and enabled at marking M. Now, if tim∈T fires, the new
marking M’ such that M[tim(rr)>M’ is as follows:
M’(zi)= M(zi)-C-(zi,tim)(<rr(v)>)=<0> (1)
+
M’(zm)=M(zm)+C (zm,tim)(<rr(v)>)=<rr’(v)> (2)
where rr’(v)=(zm, …z1) is obtained by applying the function “UP” to rr(v)=(zi zm, …z1).
Finally, the stamp s(rr(v)) is reset to zero.
<(zi , zm , ..., z 1)>
tqi tim

Id Id

UP UP
zq zi zm

Figure 2. Example of transitions in a CTPN.

9
 Example 2. Let us consider the AGVS described in Example 1, with NV=5 and the
following routes assigned to the AGVs: rr(v1)= (z7, z2, z3, z8, z7, z1), rr(v2)=(z2, z3, z8, z7, z1),
rr(v3)=(z9, z8, z10, z1), rr(v4)= (z3, z8, z7, z1), rr(v5)= (z1). We remark that v5 is a vehicle
waiting for destination in the docking station. Each time delay is supposed to be δ(zm,tim)=2
for each zm∈P and tim∈T. Figure 3 depicts the corresponding CTPN model at marking M1 and
at time τ. The marking and the time stamps are defined as follows: M1(z7)=<rr(v1)>=<(z7, z2,
z3, z8, z7, z1)>, M1(z2)=<rr(v2)>=<(z2, z3, z8, z7, z1)>, M1(z9)=<rr(v3)>=<(z9, z8, z7, z1)>,
M1(z3)=<rr(v4)>=<(z3, z8, z7, z1)>, M1(z1)=<rr(v5)>=<(z1)>, s(rr(v1))=0, s(rr(v2))=1,
s(rr(v3))=3, s(rr(v4))=2, s(rr(v5))=5. Figure 3 shows the colour and the stamp associated with
each token. Note that the token colour in z7 enables t72 and does not enable t71. However,
since M1(z2)≠<0>, transition t72 is inhibited by condition C1).

Before closing this section, we recall that two types of events can change the system state
(i.e., the marking of the CTPN) and trace the AGVS behaviour:

z4
Id <rr(v3)>, s(rr(v3))=
UP =<(z 9, z8, z 7, z1)>, 3
Id UP
z8
UP t84 t98 t49
t38 UP Id UP
t59
<rr(v4)>, s(rr(v4))= z 9 Id
=<(z 3, z8, z7, z 1)>, 2 Id UP UP
Id z5
Id Id
z3 t89
UP UP
t65
UP t87
t23 t10,9 t9,10 Id
Id t78
<rr(v2)>, s(rr(v2))= z6
=<(z 2, z3, z8, z 7, z1)>, 1 UP UP
z2 Id
Id
<rr(v5)>, s(rr(v5))=<(z1)>, 5 UP
t10,6
UP t71 t10,1
t72 Id
Id Id UP UP Id

<rr(v1)>, s(rr(v5))=<(z7, z2, z3, z8, z 7, z1)>, 0 z7 z 10

UP Id z1 Id UP

t17 t1,10

Figure 3. The CTPN at marking M1 for Example 2.

10
i) a new path r is assigned to a vehicle v∈V (1-type event). This event is identified by the pair
σ1=(v, r). Obviously, the first entry in the path r must be the zone currently occupied by v;
ii) a vehicle moves from a zone to another one (2-type event). This event is identified by the
symbol σ2=v.
The occurrence of a 1-type event σ1=(v, r) assigns the new colour r to v. Such an event can
occur when the vehicle is waiting in the docking station or when the AGV is returning to the
docking station after completing a task. On the other hand, if a 2-type event σ2=v happens
with rr(v)=(zi, zm, …z1), transition tim fires in the CTPN and the marking M is updated
according to equations (1) and (2). Let Σ1 and Σ2 respectively indicate the sets of 1- and 2-
type events.
We remark that in the AGVS behaviour some scheduling alternatives can be considered.
For example, when several transitions are simultaneously colour enabled and ready at time τ,
it is necessary to choose the AGV that has to move to the next zone first. To this aim, a
priority rule is defined by a function that associates with each marking M and each instant τ a
vehicle from V, i.e.,
π: MA×ℜ+→V (3)
where MA is the set of admissible markings of the CTPN.
The present article focuses on the choice of the deadlock avoidance policy, rather than on
the priority rule selection. However, it is important to bear in mind that the system
performance is affected by the choice of the priority rule π.

Example 3. Let us consider again the AGVS of Example 1, described by the CTPN
marking M1 and at time τ in figure 3. Since it holds M1(z9)≥C-(z9, t98)(<rr(v3)>) with
rr(v3)=(z9, z8, z7, z1) and M1(z3)≥C-(z3, t38)(<rr(v4)>) with rr(v4)=(z3, z8, z7, z1), the enabled
transitions are t98 and t38. It holds s(rr(v3))=3>δ(z9, t98)=2 and s(rr(v4))=2=δ(z3, t38), hence
both t98 and t38 are ready. Now, the priority rule π selects the AGV that has to move first to
the next zone. If we choose the “First Come First Served” (FCFS) priority rule, this selects the
following vehicle: π(M1,τ)=v3, since v3 arrived first in the zone it is holding.

11
4. Deadlock conditions and digraphs

4.1 Previous results

This paper faces the deadlock problem in AGVSs using an approach employed in the
context of deadlock avoidance in AMSs (Fanti et al. 1997) and successively modified for
application to AGVSs (Fanti 2002a). More precisely, in the present framework vehicles play
the role of jobs for acquiring resources (i.e., zones). Moreover, the system output is
represented by the zone z1 that can accommodate all the vehicles.
Here, we recall the main definitions and results necessary to explain the deadlock
avoidance strategies. The proposed techniques use digraphs to characterize deadlock. In
particular, all the current interactions between vehicles and zones can be described by means
of a digraph DT(M)=(N,ET(M)) named Transition digraph and depending on the current
CTPN marking. Each vertex in N corresponds to a zone zi, so that the same symbol is used for
vertices and zones, i.e., N=Z. The vertex set is fixed, but the edge set changes at each event
occurrence and is defined as follows: eim∈ET(M) iff there exists at marking M a vehicle v∈V
occupying zi and requiring zm as the next zone. The following proposition is proven by Fanti
et al. (1997) and, concerning the AGVSs context, in Fanti (2002a):

Proposition 1. The AGVS is in deadlock condition in the current marking M iff there
exists a cycle in the transition digraph that does not contain zone z1.

However, a deadlock avoidance policy must prevent not only the deadlock conditions
characterized by Proposition 1, but also unsafe states that are not deadlock but can incur a
deadly embrace in the next future. To define an efficient deadlock avoidance strategy, a state
named second level deadlock is characterized by the definition of a second digraph
DR(M)=(N,ER(M)) called residual path digraph (Fanti 2002a). Digraph DR(M) is built taking
into account the residual path that each vehicle in the system has to complete at the current
marking M. Hence, eim∈ER(M) iff zm immediately follows zi in rr(v), for some v∈V.
Obviously, DT(M) is a subdigraph of DR(M) and they both depend on the current state of the
system. Now, a second level deadlock can occur only if the cycles of DR(M) enjoy a
particular property that can be exhibited using a further digraph, D2R(M)=(N2(M),E2R(M)),
named second level digraph and obtained from DR(M) as follows. Denoting by {γ1, γ2, …,
γM} the complete set of the cycles of DR(M) not including z1, we associate a vertex γk∈N2(M)
to each cycle γk of DR(M). Moreover, an edge e2kh is in E2R iff the following two conditions
12
hold: a) γh and γk have only one vertex in common (say rm); b) there exists a residual path
rr(v) for some v∈V requiring zones zi, zm and zp in strict order of succession and eim is an
edge of cycle γh, while emp is an edge of cycle γk.
Now, let γ2 be a cycle from D2R(M) (second level cycle) and let Γ2(M) be the subset of
second level cycles enjoying the following property: γ2∈Γ2(M) iff the cycles associated with
the vertices of γ2 are all disjoined but for one vertex, common to all of them. Moreover, let the
capacity of a cycle γ (denoted by C(γ)) be defined as the number of resources involved in
such a cycle. Analogously, let us define the capacity of a second level cycle C(γ2) as the
number of distinct resources involved in all the cycles corresponding to the vertices of γ2.
Finally, let C20(M) be the minimum capacity of the second level cycles from Γ2(M)
(C20(M)=∞ if Γ2(M) is empty). Considering that nV(M) indicates the number of vehicles
performing transport operations in the current marking (not including the vehicles waiting in
the docking station) the following proposition, proven in Fanti (2002a), is here reformulated
with reference to the CTPN framework.

Proposition 2: A marking M can be a second level deadlock state for the AGVS only if
Γ2(M) is not empty and nV(M)≥(C20(M)-1).

4.2 Relation between CTPN and digraphs

In this section we establish the relation between the described CTPN model and the
introduced digraph tool for deadlock detection. To this aim, we remind that a digraph
containing N nodes is completely characterized by its (N×N) adjacency matrix (Harary 1971).
We assume that the (j,i)-th entry of the adjacency matrix is equal to one iff the digraph
contains an arc from the i-th node to the j-th one.
Modelling the AGVS by the CTPN, it is possible to obtain the adjacency matrix of DT(M)
at each marking M using the incidence matrix of the CTPN. Since the transition digraph
depends on the marking of the CTPN, to build the adjacency matrix AM of DT(M) we define
the (|P|×|T|) matrices IM and O as follows:
IM(zi,tij)=1 if there is a coloured token in zi∈P enabling tij∈T, otherwise IM(zi,tij)=0.
O(zi,tim)=1 if C-(zi,tim)≠0, else O(zi,tim)=0.
Now, matrices AM and IM depend on the marking M of the CTPN and can be built by the
following algorithm.

13
Matrices algorithm
Step 1 Set i=1
Step 2 If M(zi)≠<0> and zj is the second resource of <rr>=M(zi), then IM(zi,tij)=1
Step 3 If i<|P| then i=i+1, go to step 2
Step 4 AM=O·IMT (Boolean product).

According to the Matrices algorithm, the (j,i)-th entry of the (|P|×|P|)-matrix AM is one iff
there exists a coloured token in zi enabling the acquisition of zone zj. This is equivalent to
saying that eij∈ET(M) and leads us to the conclusion that AM is the adjacency matrix of the
transition digraph.

Finally, the (|P|×|P|)-adjacency matrix AR of digraph DR(M) can be obtained by the CTPN
marking as follows:
AR(zi,zj)=1 if there exists z∈P such that (zi, zj) is a subsequence of M(z)=<rr>.

Example 4. Let us consider the system described in Example 2 and let us define the priority
rule π as follows:
π(M,τ)=vk=min {s(rr(vh))⏐ s(rr(vh))≥δ(zi) with M(zi)=<rr(vh)>}. (4)
h

More precisely, the priority rule selects, among the vehicles enabling ready transitions, the
one with minimum time stamp. Even if at marking M1 transitions t72, t38, t23 and t98 in figure 3
are colour enabled, only t38 and t98 are ready at time τ, because s(rr(v3))>δ(z9, t98) and
s(rr(v4))=δ(z3, t38). Between v3 and v4 the priority rule π chooses v3. Since transition t98 is
enabled by (1) and (2), it fires, then the stamp s(rr(v3)) is set to zero and the marking M2,
such that M1[t98(<rr(v3)>)>M2, is the following: M2(z9)=<0>, M2(z8)=<rr(v3)>=<(z8, z10,
z1)>, M2(zi)=M1(zi) for each zi∈P with i≠8,9.
Figure 4 shows the CTPN at marking M2 and figure 5 shows the corresponding transition
digraph that exhibits the cycle γ=({z2, z3, z8, z7,},{e23, e38, e87, e72}) not including z1. Hence,
according to Proposition 1 the AGVS is in deadlock condition.

14
 z4
UP Id
<rr(v3)>, s(rr(v3))=<(z8, z 7, z1)>, 0 UP
Id
z8
UP t84 t98 t49
t38 UP Id UP
t59
<rr(v4)>, s(rr(v4))= z9 Id
=<(z3, z 8, z7, z1)>, 2 Id UP UP
Id z5
Id Id
z3 t89
UP UP
t65
UP t8 7
t23 t10,9 t9,10 Id
Id t78
<rr(v2)>, s(rr(v2))=
=<(z2, z 3, z8, z7, z 1)>, 1 z6
UP UP
z2 Id
Id
<rr(v5)>, s(rr(v5))=<(z 1)>, 5 UP
t10,6
UP t71 t10,1
t72 Id
Id Id UP UP Id

<rr(v1)>, s(rr(v1))=<(z 7, z2, z3, z 8, z7, z 1)>, 0 z7 z10

UP Id z1 Id UP

t17 t1,10

Figure 4. The CTPN at marking M2 for Example 4.

z4
z8 z9

v4
z5
z3
v3
v2
z2
z1 z6

v1 z7 z10

Figure 5. Digraph DT for Example 4.

5. Definition of the control policy

This section presents two control policies (CPs) to avoid deadlocks and collisions in zone-
controlled AGVSs. The first CP was introduced by Fanti (2002a) and the second one is
defined in this paper. The two policies share the same algorithm checking for zone validation
and differ for the path validation algorithm.

15
5.1 Control policy 1
In terms of Petri nets modelling, a transition tim∈T is said to be controlled if its firing is
determined by a CP when tim is enabled according to conditions C1 and C2. Therefore, a CP is
a restrictive policy that determines whether a controlled transition can fire on the basis of
knowledge of the net state. When a controlled transition tim∈T can fire according to a CP, we
say that tim is control enabled under this CP. Moreover, if at least one transition is control
enabled, the CTPN is said controlled under this CP. More formally, a CP is a mapping that
associates with each event σ∈Σ1∪Σ2 and with each marking M a control action that enables
and inhibits event σ, i.e., distinguishing between 1-type and 2-type events:

fi : Σi×MA→{0,1} with i=1,2 (5)

where fi(σi,M)=0 (fi(σi,M)=1) means that for the CTPN at marking M event σi is control
inhibited (control enabled) for i=1,2. In particular, governing by f2 a 2-type event σ2=v means
to control the corresponding transition tim, where ri and rm are the first and the second zones of
rr(v). Note that transitions ti1∈T do not require any control, because their occurrence does not
determine deadlocks and collisions.
From the CTPN point of view, a deadlock means that once some marking has been
reached, some colour enabled transitions cannot be fired any longer. The consequence in our
model is that the travel of an AGV along a path is interrupted and the vehicle can not reach
the docking station. Accordingly, we define task marking the marking associated with the
system state in which all the AGVs are in the docking station and have completed their
assigned routes. More formally, we introduce the following definition:

Definition 1: Marking M* is called task marking if M*(z1)=NV<z1> (i.e., NV tokens are in

z1 with colour <z1>) and for each zi such that zi≠z1 it holds M*(zi)=<0>.

To avoid deadlock it is sufficient to guarantee that each AGV route can continue so that all
vehicles can reach the docking station under the control policy. Accordingly, the following
definition is given:

Definition 2: A CTPN is deadlock-free at marking M∈R(M0) under a CP iff there exists a

controlled firing sequence ω so that M[ω>M*.

16
 According to Definition 2, the deadlock avoidance controller can be synthesized by finding
for the CTPN model a CP that keeps the CTPN deadlock-free, i.e., able to reach the task
marking M*.
Starting from Proposition 1, a deadlock avoidance policy can be defined in a
straightforward manner, on the basis of a look-ahead procedure including just one step (Fanti
et al. 1997, Fanti 2002a). More precisely, when an event has to occur, the policy updates the
next marking M’, builds the adjacency matrix of the new transition digraph DT(M’) and
inhibits the event iff such a digraph contains a cycle not including z1. Such a control policy is
defined as follows:

fi(σi,M)=0 if the transition digraph associated with M’ and described by the adjacency
matrix AM’ exhibits a cycle not including z1.
fi(σi,M)=1 elsewhere, with i=1,2.

As shown in (Fanti et al. 1997), this control algorithm may lead the system to a situation
that is called restricted deadlock. Although the system is not in a deadlock state, in such a
condition it inevitably incurs a permanent blocking caused by the control inhibition, therefore
the task marking M* is not reached. Since it is shown that the defined CP incurs a restricted
deadlock iff a second level deadlock occurs, the defined CP is deadlock free if the necessary
condition of Proposition 2 is violated. Hence, the following control policy, named CP1, is
deadlock and restricted deadlock free.

CP1
Let the CTPN be at time τ and at marking M and denote with M’ the marking obtained
from M after the occurrence either of σ1∈Σ1 or of σ2∈Σ2. The restriction policy is defined as
follows:
f1(σ1,M)=0 if the transition digraph associated to M’ and described by the adjacency matrix
AM’ exhibits a cycle that does not include z1 or if Γ2(M’) is not empty and
nV(M’)≥(C20(M’)-1).
f1(σ1,M)=1 elsewhere.
f2(σ2,M)=0 if the transition digraph associated to M’ and described by the adjacency matrix
AM’ exhibits a cycle that does not include z1.
f2(σ2,M)=1 elsewhere.

17
 By Proposition 2, we deduce that if the CTPN is at the initial marking M0 characterized by
a transition digraph that does not exhibit a cycle including z1 and if nV(M0)<(C20(M0)-1), then
the CTPN behaviour is deadlock free under CP1. Indeed, since in such a case a deadlock and
a second level deadlock can not occur, there exists a controlled firing sequence ω so that
M[ω>M*. The recalled control policy is the less restrictive policy proposed in Fanti (2002a).
We remark that such a policy suffers from high off-line computational complexity. A further
disadvantage of CP1 is that the policy permissiveness is considerably diminished by the
presence of bi-directional guide-paths in the AGVS layout.

5.2 Control policy 2

To overcome the drawbacks of CP1, we introduce a second control policy based on an
analysis of the task marking reachability.

Definition 3: Let the CTPN be at marking M’∈R(M0) and at time τ, and let us suppose
that f1(σ1,M)=0 for each σ1∈Σ1 and M∈MA. The task marking M* is said to be reachable
from M’ under f2 and the priority rule π, if there exists a controlled firing sequence ω so that
M’[ω>M*.

In the following we introduce the novel control policy CP2.

CP2
Let the CTPN be at time τ and at marking M and let M’ denote the marking obtained from
M after the occurrence either of σ1 or of σ2. The restriction policy is defined as follows:
f1(σ1,M)=1 if M*is reachable from M under f2 and the priority rule π.
f1(σ1,M)=0 elsewhere.
f2(σ2,M)=0 if the transition digraph associated to M’ and described by the adjacency
matrix AM’ exhibits a cycle that does not include z1.
f2(σ2,M)=1 elsewhere.

The following proposition proves that the system controlled by CP2 is deadlock and
restricted deadlock free.

18
 Proposition 3: Let the CTPN be at marking M∈R(M0) at time τ. If M* is reachable from
M under f2 and the priority rule π, the CTPN is deadlock free under CP2.

Proof
Let us suppose that the CTPN is at time τ and at marking M∈R(M0). If we assume that no 1-
type event occurs, by applying f2 and a priority rule π only one possible firing sequence ω can
be obtained to reach marking M*. By assumption such a sequence exists and successively f2
controls the 2-type events inhibiting transitions that determine deadlock at the next step.
Moreover, f1 enables a 1-type event if there exists a controlled firing sequence ω under f2 and
the priority rule π, so that the CTPN can reach M*. So the AGVS performance will follow the
event sequence ω as long as a new 1-type event occurs. Hence, the system deadlock freeness
is guaranteed.

6. Synthesis of the controller

This section describes the two activities of the real time controller: the zone validation and
the path validation that implement CP1 and CP2. While the zone validation algorithm is the
same for the two policies, CP1 and CP2 differ for the PVA.

6.1 Zone validation

Let us suppose that the CTPN is at marking M and that transition tim is colour enabled, i.e.,
M(zi)=<rr(v)>=<(zi, zm, …z1)>. If s(rr(v))≥ δ(zi) holds, event σ2=v can occur. In such a case,
the controller executes the following Zone Validation Algorithm (ZVA).

ZVA
A1 If M(zm)≠<(0)> by C1) the transition cannot fire, the zone is not validated. Go to step A5.
A2 If M(zm)=<(0)> then the controller determines the new marking:
M’(zi)=<(0)>
M’(zm)=<(zm, …z1) >
M’(z)=M(z) for each z∈P with z≠zm, and z≠zi.
A3 The Matrices algorithm provides matrices IM and AM’.
A4 A depth-first search algorithm (Reingold et al. 1977) is applied to AM’:
if the search finds a cycle not including z1 then f2(σ2,M)=0 and go to step A5,
19
 else f2(σ2,M)=1 and go to step A6.
A5 The zone is not validated, the behaviour of the AGVS continues with the CTPN at
marking M. STOP.
A6 The zone is validated and tim fires, the behaviour of the AGVS continues with the CTPN
at marking M’. STOP.

We remark that Step A1 allows us to avoid collisions and step A4 tests whether marking
M’ is a deadlock state or not.

6.2 Path validation

The path validation consists in enabling or inhibiting 1-type events in order to avoid
immediate deadlocks and restricted deadlocks as long as a new path is assigned to some
vehicle. Let r be the path that the scheduler proposes for vehicle v and let M be the current
marking of the CTPN at time τ. Under CP1, the controller executes the following Path
Validation Algorithm (PVA).

PVA1
B1 Update the marking of the CTPN as follows:
M’(zk)=<r(v)> where zk is the first resource of r(v)
M’(zi)=M(zi) for each zi∈P with zi≠zk.
B2 A depth-first search algorithm is applied to AM’:
if the search finds a cycle not including z1 then f1(σ1,M)=0 and go to step B4.
B3 Build the path digraph and the second level digraph. If Γ2(M’) is not empty and
nV(M’)≥(C20(M’)-1), then f1(σ1,M)=0 and go to step B4, else go to step B5.
B4 r(v) is not validated and the evolution of the CTPN continues starting at marking M.
STOP.
B5 r(v) is validated and the evolution of the CTPN continues starting at marking M'. STOP.

Under CP2, the real time controller is to validate the proposed route by using PVA2 which
works according to the following steps.

PVA2
B1 Update the marking of the CTPN as follows:

20
 M’(zk)=<r(v)> where zk is the first resource of r(v)
M’(zi)=M(zi) for each zi∈P with zi≠zk.
B2 Start a simulation run that under the ZVA and a priority rule π checks that the CTPN
reaches marking M* from M’.
B3 If marking M* is reached then r(v) is validated and the evolution of the CTPN continues
starting at marking M’, else r(v) is not validated and the evolution of the CTPN continues
starting at marking M. STOP.

6.3 Some remarks about the computational complexity of the proposed strategies
The zone validation activity exhibits the complexity of the depth-first search algorithm
(Reingold et al. 1977), which is polynomial with complexity O(NZ).
As shown in Fanti (2002a), PVA1 has a polynomial complexity with the number of cycles
2
D R(M) and DR(M). Since these numbers grow exponentially with the number of nodes of the
digraphs, PVA1 may exhibit high complexity. In fact, each bi-directional guide-path
corresponds to a first level cycle and each couple of bi-directional guide-paths sharing a node
in the digraph representing the system corresponds to a second level cycle (Fanti 2002a).
Hence, with reference to figure 1, if all the guide paths in the case study were bi-directional,
CP1 would exhibit a high computational complexity, while being severely restrictive. More
precisely, in such a case the number of cycles in the first level and second level digraphs
would be c1=(NZ-1)!=9! and c2=((NZ-1)!-1)!=(9!-1)! respectively, hence the path validation
under PVA1 would require O(9!(9!-1)!) operations (Fanti 2002a). On the other hand, the
computational complexity of the proposed control policy CP2 does not depend on the number
of cycles in the system layout. In particular, implementing PVA2 requires a simulation run
where NVL events can occur (L being the maximum length of each assigned route). For each
2-type event of the simulation, a zone validation must be verified with complexity O(NZ).
Moreover, for each 2-type event at most NV zone validations verify whether an inhibited
event can be enabled. Summing up, the path validation under PVA2 requires O(NV2LNZ)
operations. Hence, with reference to the layout in figure 1, regardless of the number of bi-
directional guide-paths, it holds NZ=10 and assuming NV=5 and L=2NZ=20 we get 5·103
operations.

21
7. Simulation study
This section presents a simulation study performed for the AGVS described in Example 1
employing in turn the control policies CP1 and CP2. The aim of the present investigation is to
compare the effectiveness of CP2 with CP1 by means of appropriate performance measures.
Additionally, in the simulations the AGV fleet size is varied, in order to check the control
policy effectiveness when traffic changes from low to medium and finally becomes intense.
The CTPN representing the system is implemented in the MATLAB-Stateflow
environment (The MathWorks Inc. 1997), where it is possible to integrate modelling and
simulation of Stateflow event-driven systems (e.g., the AGVS dynamics) with the execution
of MATLAB computation routines (e.g., developing the ZVA and PVA algorithms), while
keeping track of time by way of a software clock.
The rationale for choosing MATLAB-Stateflow as simulation environment, rather than
more evolved and specific software, e.g. such as Arena (Kelton et al. 1998), lies in its
simplicity and immediacy. In fact, rather than mimicking the operation of an actual AMS in
its complexity (workstations, buffers working procedures, etc.), the purpose of the following
simulation study is testing the efficiency of several traffic control policies. Hence, we
investigate, for equivalent settings, which policy restricts the system congestion while
guaranteeing that the vehicles moving in the system reach their destination at an early time.
Summing up, our purpose is to simulate the CTPN while assigning the vehicles the longest
and most varied paths. To this aim, MATLAB exhibits two convenient features: 1) simplicity
of implementation and ability to reproduce the CTPN modularity and structure; 2) possibility
to execute, when required, a simulation run implementing PVA2. On the contrary, using a
software environment such as Arena to reproduce the CTPN would require the
implementation of computationally intensive cycles, e.g. in C++ code, in order to execute a
simulation run during the system evolution.
A detailed discussion of the MATLAB-Stateflow AGVS model is beyond the scope of this
paper, nevertheless in the sequel we briefly describe the model rationale, since the execution
of the CTPN dynamics in MATLAB-Stateflow is carried out by a novel implementation
technique. More precisely, here the Petri net places are represented by a finite state
automaton, with sub-states modelling the zone shift of the AGV currently under investigation
by the AGVS management system. For sake of brevity, we do not show the details of such
technique. However, figures 6 and 7 show the MATLAB-Stateflow models implementing the
AGVS controlled by restriction policies CP1 and CP2, respectively. Both charts in figures 6
and 7 comprise three main subsystems: “AGVS Dynamics”, “Buffer” and “PVA”. Note that
22
the first two blocks are nearly the same in both machines, on the contrary the last block differs
since it implements in turn PVA1 and PVA2. More precisely, the “AGVS Dynamics” block
handles the vehicle travels in the system and it is composed of two procedures: “Place” and
“Vehicle Moves and ZVA”. Now, the first procedure describes the generic CTPN place, i.e., a
generic zone; the second one is in charge of selecting the next vehicle to move and executing
ZVA for the selected AGV. Moreover, the “AGVS Dynamics” block updates and stores the
current marking of the CTPN. When a new route is proposed, the “Buffer” block is invoked.
Indeed, it saves the current CTPN marking while PVA validates a new path proposed by the
“Vehicle Moves and ZVA” block, transferring the system data to the PVA block in order to
validate the new route. In addition, the “Buffer” block restores the system status in case of
path rejection, hands the target marking over in case of route authorization and informs the
“AGVS Dynamics” block about the path validation outcomes. Finally, the “PVA” block
performs the path validation and notifies its decisions to the “Buffer” block.
As already mentioned, the PVA implementation differs for the two restriction policies
considered. In particular, the “PVA1” block (see figure 6) tests the occurrence of a restricted
deadlock by checking the presence of a particular cycle in the second level digraph
originating from the new marking, according to Proposition 2. Instead, in the implementation
of PVA2 (see figure 7) the “PVA2” block tests whether Proposition 3 holds by performing a
simulation of the AGVS in order to check the task marking reaching under the ZVA control
and supposing that no 1-type event occurs. This test is accomplished by an additional “AGVS
Dynamics” sub-block that simulates the AGVS behaviour originating from the marking under
consideration.
To sum up, we remark that using the MATLAB-Stateflow environment to implement the
AGVS under study results in a compact and modular model, in which it is easy to modify the
PVA implementation. Indeed, the system model is always the one represented in figure 6
(figure 7) when PVA1 (PVA2) is employed, no matter how many zones the AGVS comprises.
In fact, in the automaton implementing the system a single state represents the generic CTPN
place and few changes are necessary to update the system layout and the possible routes.

23
Figure 6. The MATLAB-Stateflow machine implementing the CTPN under CP1.

Figure 7. The MATLAB-Stateflow machine implementing the CTPN under CP2.

24
7.1 Description of the simulation experiments
Several simulation experiments are conducted for Example 1 under the following
assumptions, employing in turn CP1 and CP2.
A set of 300 routes resulting in three replications of a set of 100 paths, all starting and
terminating with the docking station z1, is considered for the AGVS to accomplish in a
predetermined order. The maximum route length is set equal to ten, i.e., each route proposed
by the path scheduler consists of ten zones, all ending with z1.
The assigned set of 300 routes is considered in six experiments for both control policies
with an AGV fleet size NV ranging from 2 to 7.
In order to keep into account the time necessary for an AGV to travel from a zone to an
adjacent one, i.e., the time for a vehicle to move along a guide-path or accomplish a load
(unload) operation from (to) a workstation or the corresponding buffer, we set a time for a
vehicle to accomplish such zone crossing. For sake of simplicity, we assume that any AGV
can travel between any couple of adjacent zones in 200 time units, i.e., in a time interval equal
to the 500th part of the total run time.
The priority rule selected for the simulation experiments is the FCFS policy formally
defined by (4).
Since simulation experiments employing a different number of vehicles and a different
restriction policy may not require the same run time to accomplish the 300 selected missions,
a fixed run period of T=100,000 time units is considered. Finally, in the present case study we
assume that resources are reliable. Nevertheless, we remark that if a truck breakdown occurs,
the failed vehicle is removed from the AGVS layout and the system dynamics proceeds with
its evolution using a lower fleet size and by assigning a new marking to the CTPN.

7.2 Performance measures

In order to compare the effectiveness of the two control strategies in assigning the
transport tasks to the AGVS, as well as to facilitate selection of the AGV fleet size, we
consider several performance measures for the system.
For each simulation experiment, we select as measure of performance the number of routes
completed by the AGV fleet under the ZVA and PVA verifications (throughput) in the run
time T. Clearly, the number of completed missions is an index of the traffic management
policy effectiveness in having the system accomplish paths while minimizing blockings. Note
that only completed missions are considered in the throughput computation: at completion of
the run time unfinished tasks are discarded.
25
 In addition, vehicles utilization is evaluated in the simulations. The operative conditions in
which the generic vehicle activity can be classified are as follows.
(i) Loaded and booked travel – the vehicle is accomplishing a mission. This state
comprises two sub-states: the AGV is either booked and travelling empty toward the
loading zone, or transporting a piece toward the assigned unloading area.
(ii) Blocked – the AGV, while travelling loaded or booked, is blocked in a zone by the
control system in order to avoid deadlock or collision.
(iii) Empty and available – the vehicle, after unloading a part at the destination zone, is
either empty and travelling towards the docking station or idle in the docking station
waiting for destination. At both stages the AGV can receive the next transportation
task by the path scheduler.
The measure of performance adopted for vehicles activity is the average percentage time a
vehicle spends in each of the above three states. Additionally, we consider vehicles
utilization, computed as the average percentage time a vehicle spends executing a mission,
i.e., being in states (i) or (ii).

7.3 Simulation results

Results of the simulation experiments are summarized in figures 8-12. In particular, the
AGVS throughput is depicted in figure 8 and reported in detail in tables 1 and 2 for CP1 and
CP2, respectively, when using different AGV fleet sizes. Figure 8 shows that the AGVS
throughput reaches the maximum value under CP1 when five vehicles are available, whereas
the highest number of accomplished transportation tasks is obtained under CP2 with six
vehicles. Moreover, it can be inferred from figure 8 that under both PVA strategies the
transportation system is inefficient for a modest AGV fleet size, since the number of vehicles
is small compared both with the AMS size and the large number of tasks to be completed in
the assigned run time. All the same, under both real time control strategies the homogeneous
increase of the allocated missions (see figure 8) is an obvious consequence of the growth of
the available AGVs number (Rajotia et al. 1998b). However, the throughput value reaches a
maximum for a given fleet size, since traffic becomes congested and vehicles are often
blocked in order to avoid deadlock and collisions. For all AGV fleet sizes CP2 provides better
throughput values than CP1.
The AGVS throughput is reported in detail under CP1 (see table 1) and CP2 (see table 2).
Each column in tables 1 and 2 shows the simulation results under the selected CP for a
specific fleet size NV ranging from 2 to 7, reporting the number of transportation missions
26
accomplished by each vehicle and the tasks completed by the whole AGV fleet in the run
time. Comparing the corresponding columns in table 1 and table 2, i.e., the results of the
simulation runs with the same fleet size, the improvement in the overall number of completed
transportation tasks under CP2 is apparent. Clearly, CP2 allows a larger number of completed
missions than CP1, because PVA2 is less restrictive than PVA1. More precisely, PVA1
inhibits some paths because it detects the potential occurrence of restricted deadlock in the
future. On the contrary, in some of these cases PVA2 shows by simulation that the mission
can be completed if the AGVS follows a particular events sequence.
The average percentage time of loaded and booked travel under both CPs is reported in
figure 9. Again, CP2 leads to better results, exhibiting for any number of vehicles a
performance index that is always higher than the one computed with the alternative restriction
policy. However, under both techniques such an index decreases with the increase in AGV
fleet size: this is expected, because of the traffic congestion deriving from the large number of
travelling vehicles in the AGVS.
Figure 10 depicts the average percentage blocked time of vehicles. Again, it is observed
that CP2 outperforms CP1. Moreover, for both techniques the performance index increases
with the fleet size. Indeed, the probability that a vehicle is blocked while executing a
transportation task becomes higher when the system is overcrowded.
The average empty and available time is reported in figure 11. Such an index is only partly
dependant on the adopted control policy: in fact, it is greatly influenced by the selected
dispatch policy and by the transport request rate. As previously mentioned, investigating the
path scheduler is not the purpose of the present paper, therefore we assume that the issue of
assigning a proper succession of missions to the vehicles has been addressed. Despite such
dependence of this index on the dispatch policy, we can observe that CP2 outperforms CP1
again. In fact, the path validation algorithm based on the task marking reaching tends to reject
a lower number of tasks than the algorithm based on cycles detection in the digraphs
associated to the AGVS.
Finally, the AGVS utilization index is reported in figure 12. Here, the average percentage
time a vehicle is either booked travelling toward a loading station, or loaded and blocked, or
else carrying a piece towards its destination is reported. This performance index is high for
both techniques, since most of the time the vehicles are busy accomplishing a mission out of
the docking station and a task may be assigned to an AGV as soon as it finishes a mission and
starts travelling to the docking station. Nevertheless, CP2 outperforms CP1 again for any
number of vehicles in the system, in accordance with the results previously described. In other
27
words, the path validation controller of CP2 is able to validate a larger number of missions
than CP1, so that the vehicles are utilized most of the time.
As a summary of the adopted control policies performance, the distribution of the vehicles
activity time is depicted in figures 13 and 14 under CP1 and CP2, respectively. For both
techniques we observe the same time indices trends with respect to a variation in the AGV
fleet size. However, the improvement resulting from the use of CP2 is apparent from a
comparison of figures 13 and 14. Under CP2 for any number of vehicles in the system the
average engaged travel percentage time is greater than the other time indices (see figure 14).
On the contrary, under CP1 the decrease rate of the average engaged travel percentage time is
higher than under CP2, so that the index becomes lower than the average time spent by the
vehicles in the blocking states when the system is congested (see figure 13). Accordingly,
under CP2 the system is less prone to blockings than under CP1. For instance, if we consider
the simulation run with six vehicles when adopting CP2, approximately 59.5% of the overall
run time is employed by the AGVs travelling busy while accomplishing a task. The average
percentage time between accomplishment of a mission and assignment of the next task to the
same vehicle equals a satisfactory 4.9%, confirming the policy efficiency. The remaining
35.6% time is occupied by blockings, therefore vehicles are often stalled in order to avoid
deadlocks or collisions. Clearly, such a considerable value of the last index partly depends on
the adoption of the zone control methodology when managing the vehicles dynamics. On the
other hand, it can be remarked that the high system throughput (265 tasks assigned out of the
available 300, see table 1) in the total run time balances such an inconvenience.

7.4 Some remarks on the applicability of the proposed traffic control technique
We showed that the CP2 control strategy outperforms CP1 both in terms of throughput and
vehicles utilization. Nevertheless, in order to apply CP2, it is crucial to test its effectiveness in
real time. In other words, all moves and mission acceptances or refusals must be executed on
line, i.e., in a response time that can be neglected. More precisely, all decisions performed by
the real time controller are classified either as zone validations or as path validations. The first
type of decisions are executed by ZVA by detecting cycles in the transitions digraph and can
be easily implemented on line by way of a depth-first search algorithm in polynomial time
(see section 6.3). Now, because of the high computational complexity of PVA1, its
application can be impracticable when the system layout is complex. On the contrary, PVA2
takes decisions executing a simulation testing the task marking reaching and such a check is
performed in polynomial time (see section 6.3).
28
 Two issues must be enlightened. First, it is important for the implementation of PVA2 that
the controller keeps memory of the succession of the 2-type events as obtained by the
simulation run: hence, when a route is validated, the AGVS dynamics must follow the
behaviour forecasted by the simulation till a new route is validated. Second, the simulation
results and conclusions hold both under deterministic and stochastic transition times. In any
case, it is crucial that the controller stores the succession of events occurred during the
simulation run, so that the AGVS follows the forecasted events sequence.
In the present work we focus on the deadlock avoidance policy performance, hence we
make use of the MATLAB interpreted language in order to perform a simulation run (i.e., the
PVA2 implementation) in a simulated environment. However, the possible implementation of
the actual controller in an industrial setting may be simply and more rapidly carried out by
executing a custom code on a central processing unit connected to appropriate sensors. In fact,
it is essential for the on-line applicability of the proposed deadlock avoidance policy that such
a simulation is performed in real time. In other words, after accomplishing the PVA2
algorithm, the AGVS must still be in the state previous to the path validation, otherwise the
simulation does not mirror the authentic system dynamics and the corresponding decisions
apply to a system state preceding the actual one.
As an example, consider the complex network representation of the Delphi-Harrison layout
reported by Oboth et al. (1999), with NV=7 vehicles and NZ=219 zones. Assuming that the
maximum possible length of each assigned route is L=2NZ, for such a system PVA2 requires
the execution of about 5·106 operations in real time (see section 6.3). Hence, a commercial
personal computer featuring a 2 GHz CPU performs such a number of elementary operations
in much less than 1 second. Now, a commercial AGV may reach the speed of 1 m s-1, hence
the time required by the decision does not slow down the system operation.

29
 280
CP1
260 CP2
No. of Routes Completed

240

220

200

180

160

140
2 3 4 5 6 7
AGV Fleet Size

Figure 8. Throughput for various AGV fleet sizes.

2 AGVs 3 AGVs 4 AGVs 5 AGVs 6 AGVs 7 AGVs

AGV No. of No. of No. of No. of No. of No. of
no. tasks tasks tasks tasks tasks tasks
1 79 71 51 41 41 33
2 69 55 54 44 28 26
3 - 59 52 44 33 22
4 - - 48 47 31 28
5 - - - 34 33 24
6 - - - - 31 30
7 - - - - - 36
TOTAL 148 185 205 210 197 199

Table 1. Detailed results of AGVS throughput for various AGV fleet sizes under CP1.

2 AGVs 3 AGVs 4 AGVs 5 AGVs 6 AGVs 7 AGVs

AGV No. of No. of No. of No. of No. of No. of
no. tasks tasks tasks tasks tasks tasks
1 72 67 65 56 47 36
2 76 64 49 52 44 33
3 - 67 61 51 41 37
4 - - 60 56 46 36
5 - - - 48 48 42
6 - - - - 39 32
7 - - - - - 41
TOTAL 148 198 235 263 265 257

Table 2. Detailed results of AGVS throughput for various AGV fleet sizes under CP2.
30
 110,0
Average Time AGVs Loaded and Booked Travel [%]

CP1
100,0
CP2

90,0

80,0

70,0

60,0

50,0

40,0

30,0
2 3 4 5 6 7
AGV Fleet Size

Figure 9. Average time vehicle loaded and booked travel for various AGV fleet sizes.

56,0
CP1
51,0
CP2
Average Time AGVs Blocking [%]

46,0

41,0

36,0

31,0

26,0

21,0

16,0

11,0

6,0
2 3 4 5 6 7
AGV Fleet Size

Figure 10. Average time vehicle blocking for various AGV fleet sizes.

31
 8,0
CP1
Average Time AGVs Empty and Available [%]

7,0 CP2

6,0

5,0

4,0

3,0

2,0

1,0
2 3 4 5 6 7
AGV Fleet Size

Figure 11. Average time vehicle empty and available for various AGV fleet sizes.

100,0
CP1
99,0
CP2
Average Time AGVs Utilization [%]

98,0

97,0

96,0

95,0

94,0

93,0

92,0

91,0

90,0
2 3 4 5 6 7
AGV Fleet Size

Figure 12. Average time vehicle utilization for various AGV fleet sizes.

32
 100,0
Loaded and Booked Travel
90,0
Blocked
80,0
Empty and Available
70,0
Average Time [%]

60,0

50,0

40,0

30,0

20,0

10,0

0,0
2 3 4 5 6 7
AGV Fleet Size

Figure 13. Vehicle activity average time under PVA1.

100,0
Loaded and Booked Travel
90,0
Blocked
80,0
Empty and Available
70,0
Average Time [%]

60,0

50,0

40,0

30,0

20,0

10,0

0,0
2 3 4 5 6 7
AGV Fleet Size

Figure 14. Vehicle activity average time under PVA2.

33
8 Conclusions
This paper proposes a Coloured Timed Petri Net (CTPN) model and an efficient control
strategy to avoid deadlock and collisions in zone-controlled Automated Guided Vehicle
Systems (AGVSs). The peculiarity of the resulting model consists in its modularity, so that
the structure of the CTPN can be automatically built and easily updated if the system layout
changes. Moreover, the assigned paths, i.e., the token colours, are modified as the vehicles
move from zone to zone or a new route is assigned. In addition, the proposed control strategy
can be easily applied and improves both the permissiveness in zone acquisition and the
computational complexity of the traffic management strategies proposed by one of the authors
(Fanti 2002a). More precisely, we modify the Path Validation Algorithm (PVA) that checks
the assignment of the new missions to vehicles in order to guarantee the success of all the
AGV assigned travels. Such a validation is performed in polynomial time by a simulation run
forecasting the AGVS behaviour. Several simulation experiments performed for an AGVS
with varying fleet size show the effectiveness of the proposed control strategy compared to a
policy previously suggested. To this aim, we propose a MATLAB-Stateflow environment that
is suitable to reproduce the modularity of the CTPN and can be an efficient tool to implement
the proposed PVA in the AGVS shop floor. Future research has to investigate how the
integration of the proposed deadlock avoidance strategy with the existent scheduling policies
affects the AGVS performance.

REFERENCES
ASHAYERI, J., GELDERS, L. F., and VAN LOOY, P. M., 1985, Micro- computer simulation in design of automated guided vehicle
systems. Material Flow, 2(1), 37-48.
BANASZAK, Z. A., and KROGH, B. H., 1990, Deadlock Avoidance in Flexible Manufacturing Systems with Concurrently
Competing Process Flows. IEEE Transactions on Robotics and Automation, 6(6), 724-734.
BLAIR, E. L., CHARNSETHIKUL, P., and VASQUES, A., 1987, Optimal routing of driverless vehicles in a flexible material
handling system. Material Flow, 4(1+2), 73-83.
CHO, H., KUMARAN, T. K., and WYSK, R. A., 1995, Graph-Theoretic Deadlock Detection and Resolution for Flexible
Manufacturing Systems. IEEE Transactions on Robotics and Automation, 11(3), 413-421.
CHU, F., and XIE, X., 1997, Deadlock analysis of Petri nets using siphons and mathematical programming. IEEE Transactions
on Robotics and Automation, 13(6), 793-804.
CO, G.C., and TANCHOCO, J.M.A., 1991, A review of research on AGVS vehicle management. Engineering Costs and
Production Economics, 21, 35-42.
DESROCHERS, A. A., and AL-JAAR, R. Y., 1995, Applications of Petri Nets in Manufacturing Systems. (New York: IEEE
Press).
EGBELU, P. J., and TANCHOCO, J. M. A., 1984, Characterization of automatic guided vehicle dispatching rules. International
Journal of Production Research, 22(3), 359-374.
34
EGBELU, P. J., and TANCHOCO, J. M. A., 1986, Potentials for bi-directional guide-path for automated guided vehicle based
systems. International Journal of Production Research, 24(5), 1075-1097.
EZPELETA, J., COLOM, J. M., and MARTINEZ, J., 1995, A Petri Net Based Deadlock Prevention Policy for Flexible
Manufacturing Systems. IEEE Transactions on Robotics and Automation, 11(2), 173-184.
EZPELETA, J., and COLOM, J. M., 1997, Automatic synthesis of colored Petri net for the control of FMS. IEEE Transactions on
Robotics and Automation, 13(2), 327-337.
FANTI, M. P., MAIONE, B., MASCOLO, S., and TURCHIANO, B., 1997, Event Based Feedback Control for Deadlock Avoidance in
Flexible Production Systems. IEEE Transactions on Robotics and Automation, 13(3), 347-363.
FANTI, M. P., MAIONE, B., and TURCHIANO, B., 2000, Comparing Digraph and Petri Net Approaches to Deadlock Avoidance
in FMS. IEEE Transactions on Systems, Man and Cybernetics. Part B: Cybernetics, 30(5), 783-798.
FANTI, M. P., MAIONE, G., and TURCHIANO, B., 2001, Distributed event-control for deadlock avoidance in automated
manufacturing systems. International Journal of Production Research, 39(9), 1993-2021.
FANTI, M. P., 2002a, Event-based controller to avoid deadlock and collisions in zone control AGVS. International Journal of
Production Research, 40(6), 1453-1478.
FANTI, M. P., 2002b, Deadlock-Free Control in Automated Guided Vehicle Systems. In P. Ezhilchelvan and A. Romanovsky
(eds.), Concurrency in Dependable Computing (The Netherlands: Kluwer Academic Publishers).
FELDMANN, K., and COLOMBO, A.W., 1998, Material Flow and Control Sequence Specification of Flexible Systems Using
Coloured Petri Nets. International Journal of Advanced Manufacturing Technology, 14(10), 760-774.
HARARY, F., 1971, Graph Theory, (Reading: Addison-Wesley Publishing Company).
HSIEH, F.S., and CHANG, S. C., 1994, Dispatching-Driven Deadlock Avoidance Controller Synthesis for Flexible
Manufacturing System. IEEE Transactions on Robotics and Automation, 10(2), 196-209.
HSIEH, S., 198, Synthesis of AGVS by Coloured-Timed Petri Nets. International Journal of Computer Integrated
Manufacturing, 11(4), 334-346.
HUANG, Y., JENG, M., XIE, X., and CHANG, S., 2001, Deadlock prevention policy based on Petri nets and siphons.
International Journal of Production Research, 39(2), 283-305.
JENSEN, K., 1992, Colored Petri Nets, Basic Concepts, Analysis methods and Practical Use, Vol I, EATS Monography and
Theoretical Computer Science (New York: Springer Verlag).
JIANG, Z., ZUO, M. J., FUNG, R. Y. K., and TU, P. Y., 2000, Temporized coloured Petri nets with changeable structure (CPN-
CS) for performance modelling of dynamic production systems. International Journal of Production Research, 38(8),
1917-1945.
KELTON, W. D., SADOWSKI, R. P., and SADOWSKI, D. A., 1998, Simulation with Arena. (Boston: Mc Graw-Hill).
KIM, C. W., and TANCHOCO, J. M. A., 1991, Conflict-free shortest bi-directional AGV routing. International Journal of
Production Research, 29(12), 2377-2391.
KUMARAN, T. K., CHANG, W., CHO, H., and WYSK, R. A., 1994, A structured approach to deadlock detection, avoidance and
resolution in flexible manufacturing systems. International Journal of Production Research, 32(10), 2361-2379.
LAWLEY, M. A., REVELIOTIS, S. A., and FERREIRA, P. M., 1998, A correct and scalable deadlock avoidance policy for flexible
manufacturing systems. IEEE Transactions on Robotics and Automation, 14(5), 796-809.
LAWLEY, M. A., 2000, Integrating flexible routing and algebraic deadlock avoidance policies in automated manufacturing
systems. International Journal of Production Research, 38(13), 2931-2950.
LEE, C. C., and LIN, J. T., 1995, Deadlock prediction and avoidance based on Petri nets for zone-control automated guided
vehicle systems. International Journal of Production Research, 33, 3249-3265.
MOORE, K. E., and GUPTA, S. M., 1996, Petri Net Models of Flexible and Automated Manufacturing Systems: a Survey.
International Journal of Production Research, 34(11), 3001-3035.
NANDULA, M., and DUTTA, S. P., 2000, Performance evaluation of an auction-based manufacturing system using coloured
Petri nets. International Journal of Production Research, 38(9), 2155-2171.

35
OBOTH, C., BATTA, R., and KARWAN, M., 1999, Dynamic Conflict-free routing of automated guided vehicles. International
Journal of Production Research, 37(9), 2003-2030.
RAJOTIA, S., SHANKER, K., and BATRA, J. L., 1998, A semi-dynamic time window constrained routeing strategy in an AGV
system. International Journal of Production Research, 36(1), 35-50.
RAJOTIA, S., SHANKER, K., and BATRA, J. L., 1998, An heuristic for configuring a mixed uni/bidirectional flow path for an
AGV system. International Journal of Production Research, 36(7), 1779-1799.
REINGOLD, E. M., NIEVERGELT, J., and DEO, N., 1977, Combinatorial Algorithms. Theory and Practice. (New Jersey: Prentice-
Hall).
REVELIOTIS, S. A., and FERREIRA, P. M., 1996, Deadlock Avoidance Policies for Automated Manufacturing Cells, IEEE
Transactions on Robotics and Automation, 12(6), 845-857.
REVELIOTIS, S. A., LAWLEY, M. A., and FERREIRA, P. M., 1997, Polynomial Complexity deadlock Avoidance Policies for
Sequential Resource Allocation Systems. IEEE Transactions on Automatic Control, 42(10), 1344-1357.
REVELIOTIS, S. A., 2000, Conflict resolution in AGV Systems. IIE Transactions, 32, 647-659.
THE MATHWORKS INC., 1997, Stateflow User’s Guide Version 4.
VISWANADHAM, N., NARAHARI, Y., and JOHNSON, T. L., 1990, Deadlock avoidance in flexible manufacturing systems using
petri net models. IEEE Transactions on Robotics & Automation, 6(6), 713-722.
VISWANADHAM, N., and NARAHARI, Y., 1992, Performance Modeling of Automated Manufacturing Systems. (Englewood
Cliff, NJ: Prentice Hall).
WALKER, S. P., PREMI, S. K., BESANT, C. B., and BROADBENT, A. J., 1985, The Imperial College free-ranging AGV (ICAGV)
and scheduling system. S.E. Andersson (ed.), Proceedings of the 3rd International Conference on Automated Guided
Vehicle Systems, Stockholm, Sweden, 189-198.
WU, N., and ZHOU, M.C., 2001, Avoiding deadlock and reducing starvation and blocking in automated manufacturing
systems. IEEE Transactions on Robotics and Automation, 17(5), 657-668.
WU, N., and ZENG, W., 2002, Deadlock avoidance in an automated guidance vehicle system using a coloured Petri net model.
International Journal of Production Research, 40(1), 223-238.
WYSK, R. A., YANG, N. S., and JOSHI, S., 1991, Detection of Deadlocks in Flexible Manufacturing Cells. IEEE Transactions
on Robotics and Automation, 7(6), 853-859.
XING, K. Y., HU, B. S., and CHEN, H. X., 1996, Deadlock avoidance policy for Petri net modeling of flexible manufacturing
systems with shared resources. IEEE Transactions on Automatic Control, 41(2), 289-295, 1996.
YEH, M. S. and YEH, W. C., 1998, Deadlock prediction and avoidance for zone-control AGVS. International Journal of
Production Research, 36(10), 2879-2889
ZENG L., WANG, H. P., and JIN, S., 1991, Conflict Detection of Automated Guided Vehicles. International Journal of
Production Research, 29(5), 865-879.

36