Automatic Routing System for Intelligent Warehouses

K. T. Vivaldini; J. P. M. Galdames Member, IEEE; T. B. Pasqual, R. M. Sobral; R. C. Araújo, M.

Becker Member, IEEE; and G. A. P. Caurin, Member, IEEE

or

Abstract— Automation of logistic processes is essential to In many researches, the route is calculated considering the
improve productivity and reduce costs. In this context, minimum path. Broadbent et al. [1] presented the first
intelligent warehouses are becoming a key to logistic systems concept of conflict-free and shortest-time AGV routing. The
thanks to their ability of optimizing transportation tasks and, routing procedure described employs Dijkstra’s shortest path
consequently, reducing costs. This paper initially presents
algorithm to generate a matrix, which describes the path
briefly routing systems applied on intelligent warehouses. Then,
we present the approach used to develop our router system. occupation time of vehicles.
This router system is able to solve traffic jams and collisions, Kim and Tanchoco [9] proposed an algorithm based on
generate conflict-free and optimized paths before sending the Dijkstra’s shortest-path method and the concept of time
final paths to the robotic forklifts. It also verifies the progress windows graph for dynamic routing of AGVs in a
of all tasks. When a problem occurs, the router system can bidirectional path network. They presented formulas to
change the tasks priorities, routes, etc. in order to avoid new
consider the time curves, but this was not considered in their
conflicts. In the routing simulations each vehicle executes its
tasks starting from a predefined initial pose, moving to the results.
desired position. Our algorithm is based on Dijkstra's shortest- Maza and Castagna [12-13] added a layer of real time
path and the time window approaches and it was implemented control to the method proposed by Kim and Tanchoco. In
in C language. Computer simulation tests were used to validate [12], they proposed a robust predictive method of routing
the algorithm efficiency under different working conditions. without conflicts, and in [13] they developed two algorithms
Several simulations were carried out using the Player/Stage
to control the AGV system using a predictive method when
Simulator to test the algorithms. Thanks to the simulations, we
could solve many faults and refine the algorithms before the system is subject to risks and contingences avoiding
embedding them in real robots. conflicts in real time manner. In both, the curve time is not
mentioned.
I. INTRODUCTION Möhring et al. [14] extended the approaches of Huang,
Palekar and Kapoor [18] and Kim and Tanchoco [9], and
T HE routing task may be understood as the process of
simultaneously selecting appropriate paths for the AGVs
(Automated Vehicle Guided) among different solutions. One
presented an algorithm for the problem of routing AGVs
without conflicts at the time of route computation. In the
of the goals of routing for AGVs is the minimization of preprocessing step the real-time computation for each
cargo cost. In recent years, several algorithms, distinguished request consists in the determination of the shortest path
in two categories: static [1]-[7] and dynamic [8]-[17], have with time-windows and a following readjustment of these
been proposed to solve routing problems. In the first case, time-windows, both is done in polynomial time. By goal-
static routing the route from node i to node j is determined in oriented search, they computed times appropriate for real-
advance and is always used if a load has to be transported time routing. Extending this concept, Klimm et al. [14]
from i to j. Thus, a simple assumption is to choose the route presented an efficient algorithm to cope with the problem of
with the shortest distance from i to j. However, these static congestion and detours that also avoids potential deadlock
algorithms can not adapt to changes in the system and traffic situations. The authors in [14-15] did not consider that the
conditions. In dynamic routing, the information necessary to time window for routes with curves could imply the
determine efficient routes are dynamically revealed to the estimated time. Only in [16], Gawrilow et al. presented an
decision-maker and, as a result, various routes between i and algorithm to avoid collisions, deadlocks and livelocks
j can be chosen [18]. As one can notice, the several papers already at the route computation time, considering the
found in the literature, have not take into account important physical dimensions of the AGVs to cope with complex
features in the routing, such as path maneuvers, time curves, maneuvers and certain safety aspects that imply particular
load and unload operations, etc. applications.
Ravizza [17] focused on the online control of AGVs,
presenting a heuristic approach for task assignment and a
Manuscript received February 28, 2010. This work was supported in
part by FAPESP (Processes 2008/10477-0, 2008/09755-6, and 2009/01542- dynamic polynomial-time sequential routing to guarantee
6), CAPES (Process 33002045011P8) and CNPq (Process 134214/2009-9). deadlock and conflict-free routing of the AGVs in
K. T. Vivaldini; J. P. M. Galdames; T. B. Pasqual, R. M. Sobral; R.C. undisturbed operations. The articles cited above used
Araújo, M. Becker; and G. A. P. Caurin are with the Mechatronics
Laboratory - Mechanical Engineering Department, at EESC-USP, SP Dijkstra's Algorithm to calculate the route.
13566-900 Brazil (e-mails: {kteixeira}{galdames}{becker}{gcaurin} However, the shortest route is not always the most
@sc.usp.br; and {thales.bueno}{rafael.sobral}{roberto.carlos.araujo} efficient method, ie. the number of maneuvers is larger than
@usp.br).
the number necessary for the path to be executed. shortest-path method [20] to calculate the routes of the
In this context, as a solution to this problem, this paper robotic forklift adding heuristic functions to optimize the
presents the development of a routing system that computes quantity of maneuvers, and using the method of routing with
optimal routes, reducing the amount of unnecessary time-windows [21] to ensure conflict-free routes.
maneuvers, allowing path planning and coordinating the task
execution of the transport and handling in structured II. ROUTING ALGORITHM
environments. It also avoids known and dynamic obstacles
located in the environment. This approach investigates an
efficient solution to routing AGVs, as an alternative to the
various methods developed previously in the literature. We
also propose a software architecture that considers the local
and global tasks of the robotic forklifts (e.g.: local
navigation, obstacle avoidance, and auto-localization).
Figure 1 presents this architecture.

Fig. 2. Proposed Algorithm.

The routing algorithm (Fig. 2) was based on the dynamic
programming approach, which consists in dividing the
original problem into smaller and simpler problems. This
approach was very useful to our problem, as it presents a
sequence of decisions to be taken along a time sequence
[22]. Therefore, our routing system could be divided into
Fig. 1. Overall System Architecture proposed.[25]
seven steps.

In summary, it can be explained as follows: the routing A. Step 1

system receives information about the required The routing system receives a list of the requests as input
transportation tasks from a user interface. Based on these (Table 1 presents a summary of the requested data
data, it selects the minimum quantity of robotic forklifts considered in the storage activities). Each request is a task
necessary to execute the tasks. Then, based on a topologic defined by a sequence of pairs: loading stations (origin node)
map of the environment, it calculates the routes for the and unloading stations (destination node). Then, according
selected forklifts, checking possible collisions, traffic jams, to the requests, the routing system checks the quantity of
etc. After that, it sends the routes to each robotic forklift and robotic forklifts necessary to execute the tasks, and assigns
regularly verifies the progress of all tasks. Taking into each task to a forklift.
account the global route, each forklift calculates its own TABLE I
local path necessary to reach its goal and monitors its REQUEST DATA, INPUTS AND OUTPUTS OF OUR ALGORITHM
surroundings looking for mobile or unexpected obstacles Data Input Output
during the execution of the planned path. At a certain
Request Quantities, loading Necessary number of forklifts
frequency, based on sensor data and an environment map, (Orders) data / unloading data and allocation of each request
each robotic forklift auto-localization subsystem updates its of each product in a route for the forklifts
estimated position and informs these data to the local
navigation subsystems. The local navigation subsystem Loading Location point for Routes that the forklifts
loading the pallets should execute to carry the
compares the current position with the desired one. If the
pallets
robot deviates from the route, the local navigation subsystem Unloading Location point for Routes that the forklifts
sends commands to correct its pose returning the planned delivering the pallets should execute to unload the
route. If the local navigation subsystems verify that the route pallets
has exceeded the limit runtime determined by the routing, it
communicates it to the routing system, which recalculates Several tasks can be attributed to various robotic forklifts,
the route avoiding collisions or deadlocks among other but one task cannot be attributed to several robotic forklifts.
robotic forklifts. The algorithm is based on Dijkstra's When all tasks have been designated and the quantity of
robotic forklifts necessary has been determined, these data C. Step 3
are sent to Step 2 in order to calculate the routes. In this step the quantity of maneuvers (e.g.: curves) during
B. Step 2 the path route is analyzed by heuristic functions, which
verify the possibility of optimizing them. If the maneuvers
In this Step the routing system applies a graph-based
are unnecessary, the route is optimized using again the
approach. The graph is obtained by using a topological map
Dijkstra’s Algorithm with time-window, taking into account
of the warehouse environment. The route necessary to
the costs previously established. If the resulting total
execute each task is composed of two sub-routes. Which
execution time of the task is longer than the previously
their one has its own origin and destination nodes. Then
calculated (Step 2), the final route is not changed (it means
Dijkstra’s Algorithm is applied to calculate the lowest path
that the route cannot be optimized). At the end of this step, if
(relation between distance and total cost) for each robotic
the route has been optimized, it will be conflict-free.
forklift. The path is a continuous sequence of positions and
orientations of the robotic forklift (e.g.: the intermediate D. Step 4
positions and the pre-established positions and orientations In this step, the routes are sent to each robotic forklift and
present in Fig. 3). Aiming to guarantee that collisions the routing system interacts with the other systems
between robotic forklifts will not occur, the forklift cannot embedded in the robots (navigation and auto-localization
use the same arc or node at the same time. Therefore, we systems), as shown in Fig. 4.
know when every arc is (or not) occupied by a robotic
forklift for each calculated sub-route. These data (time
intervals [ai, bi]) and the time window are saved on a list
(log file), allowing verifying possible conflicts between
paths of the robotic forklifts. When the sub-route is a task, a
value representing the time needed to load or unload the
pallets is added to the destination node time interval.

Fig. 4. Interaction of the routing system with navigation, control, and

localizations systems.

The routing system verifies each task progress regularly.

Each robotic forklift has its own sensor, auto-localization,
and local navigation sub-system running independently and
informing regularly the global path planner about the status
of each task (position and problem found or task finished).
Therefore, it is possible to minimize the impact of this local
problem in the global system that controls the intelligent
warehouse. Table 2 presents a summary of this procedure.
TABLE II
Fig. 3. Illustration of route definition. A robotic forklift will execute a REQUEST DATA, INPUTS AND OUTPUTS OF OUR ALGORITHM
task (go to Shelf A- 01, load a pallet and unload it at Charging Platform
F). After this task, it will return to its depot. The first (1) and the second Data Input Output
(2) sub-routes represent the task execution. The third one (3) represents
the robotic forklift returning to its depot. Problems Route cannot be Inform position and list of
in route concluded. the problems found
Figure 3 illustrates how algorithm generally defines a execution (unexpected collisions,
route. The origin of the first sub-route is the current robotic exceeded time) or call
operator.
forklift position. Its destination is defined as the task initial
position (loading node). The second sub-route origin is the
loading node and its destination is the unloading node. If the E. Step 5
same robotic forklift has another task, the following sub- In this step the algorithm checks the status of the tasks,
routes are defined in a similar way. If it is closing-time, or where: 0 - task finished, 1 - rerouting of the task (unexpected
the robotic forklift needs to recharge its batteries, after the collisions or exceed time), and 2 - call operator and assign
last task its final sub-route drives it directly to its depot. tasks of this forklift to others.
Thus, we always have the origin and destination nodes and If the status is equal 1 to, it is necessary to recalculate the
also know if a task is (or not) in execution. sub-route (Here, the use of sub-routes is very useful as, only
the sub-route is corrupted). Basically the algorithm verifies Therefore, it is necessary to readjust the time windows of the
which robotic forklift presents conflicts and its location (arc subsequent sub-routes by checking the list of available time
or node – see Fig. 5). The reroute process is carried out in windows. Then, the algorithm verifies again the presence of
Step 6. conflicts in the route, due to the readjustment of the time
windows. If there is any conflict, it returns to the beginning
of this step.

Fig. 7. Sub-routes of the robotic forklift #1.

G. Step 7
Finally, the algorithm verifies if there are more tasks to be
executed. If so, it returns to its beginning (step 1). The
objective of this step is to verify and validate the execution
time of the tasks attributed versus position of the robotic
forklifts.

III. INTERFACE
Fig. 5. Illustration of two examples of collision between the robotic
forklifts #1 and #2. In (a) a collision in the arc, in (b) in the node. In this work, a graphic interface allows the routing
operation to be controlled by an operator, through Qt Creator
F. Step 6 4 software [23] and a simulated virtual environment in the
Now, the corrupted sub-route is eliminated and the arc Player/Stage developed in previous works [24]. Both were
where the conflict occurred is blocked. Then, the sub-routes implemented in Linux Operating System.
of the robotic forklifts that caused the conflict are The graphic interface (Fig. 8) was developed to perform
recalculated (Fig. 6). the communication between the user and the routing system
(including all system algorithms). It also allowed the
configuration of both routing system and simulation
environment at a higher level (Fig. 9), guaranteeing the
functionality of the routing operation and returning
information concerning the process performance in a simple
and visual way.

Fig. 6. Result of the rerouting process for the example presented in

Fig. 5-b.

To recalculate this sub-route again, the algorithm returns

to steps 2 and 3, the origin node receive the position where
forklift is and the destination node is the same for this sub-
task. These new sub-routes are called conflict-free sub-
routes. If after these conflict-free sub-routes there are other
sub-routes in the forklift schedule, then it is necessary to
verify the conflict-free final time in the destination node. If it
Fig. 8. Graphic Interface
differs from the previously calculated (Step 2), then it is
necessary to readjust the time windows of this forklift The interface allows the operator to schedule processes of
checking if it may cause other conflicts. For instance, let us robotic forklifts, setting time, data, quantity of pallets
assume that the robotic forklift #1 in Fig. 5 has more sub- involved, as well as origin and destination, and the quantity
routes scheduled, after its conflict-free sub-route (Fig. 7). of mobilized forklifts. With such information, the operator
can simulate the process. In order to do this, it is possible to the robotic forklift #4 needs to occupy the same arc at the
configure the simulation environment in the most same time that robotic forklifts #5 and #6. Unfortunately, as
appropriate and realistic way, providing drivers for the we highlighted before, if the Router did not verify all sub-
available sensors and the bi-dimensional plan of the modeled routes after the first corrupted one detected, more conflicts
warehouse. Thus, the operator can simulate and evaluate may occur. In this case, a new deadlock between the robotic
with a good accuracy the time and effectiveness of the forklifts #6 (lemon) and #1 (purple). Figure 10-b presents
process without the mobilization of real robotic forklifts. the proposed algorithm. One may verify that the conflicts
identified in Fig. 10-a were eliminated, the final paths
provided by the routing algorithm were collision-free and
avoided traffic jams. It also shows in detail the optimizations
performed by the Routing System.

Fig. 9. Graphic Interface for the simulated virtual environment.

2m
IV. RESULTS
As previously mentioned, this work used a local
navigation and auto-localization systems in order to obtain
more realistic simulations. The routing system uses the A*
algorithm in the local path planning, and applies the
Extended Kalman Filter (EKF) algorithm for the auto-
localization for each robotic forklift [24][25]. The proposed
algorithm is based on Dijkstra's shortest-path method to
calculate the routes of the robotic forklift adding heuristic
functions to optimize the quantity of maneuvers, and using
the method of routing with time-windows to ensure conflict-
free routes. Using heuristic functions, it was possible to 2m
verify the number of unnecessary curves and make a route
optimization reducing the amount of maneuvers to be
performed and thus increasing the mobility of the robotic Fig. 10. Comparison between shortest path method and optimized
forklift. It is important to emphasize that the optimized route routes of the robotic forklift #3 (grey) and forklift #2(green). In (a) the
do not exceeds the total cost previously established. route marked with an X represents the maneuver that the robotic
forklifts perform using shortest path method; and in (b) the routes
In order to verify the performance of our algorithms, marked with circles represent the maneuvers that the robotic forklifts
several simulations were carried out using a virtual did using the route optimized by our algorithm.
warehouse (50x30m) and the Player/Stage Simulator. In the In this scenario, each robotic forklift calculates the
simulation 6 robotic forklifts working at the same time in trajectory to be executed considering the route sent by the
this warehouse were considered (at this moment we did not routing system (Fig. 11). Each forklift receives a safety time
consider here unknown obstacles like people walking or for the minimum path between two nodes, them navigation
other vehicles moving in the warehouse). In order to
simplify the implementation, we selected in the Player/Stage
library the Pionner mobile robot to represent our robotic
forklifts. Maximum cruiser speeds of 1m/s and 5°/s were
applied to all robots. The algorithm could solve the conflicts
and return the optimized routes. Figure10 shows the
simulations of the tasks attributed for 6 robotic forklifts.
This test would have several collisions that can be observed
in Fig. 10-a (arrow). Firstly, the deadlock between the
robotic forklifts #4 (red) and #5 (yellow) is detected. Due to Fig. 11. Illustration of route informed for local planner. The routing
this deadlock, a new deadlock also involving the robotic system sends the each two points (e. g. forklift position and
Intermediary position) the safety time that the local planner has to
forklift #6 (lemon) occurs. The collisions happened because calculate the path planning.
system can fix both the speed and trajectory to execute the [4] N. Krishnamurthy, R. Batta, M. Karwan, “Developing conflict-
freeroutes for automated guided vehicles,” Operations Research vol.
path. The local planner applies time windows to calculate 41, pp. 1077–1090, 1993.
the safe trajectory. In case of disturbances that divert the [5] F. Taghaboni-Dutta, J. M. A. Tanchoco, “Comparison of dynamic
robot from its planned trajectory, the local planner tries to routeing techniques for automated guided vehicle system,”
International Journal of Production Research vol. 33, no. 10, pp.
run the route within the stipulated time, e.g., increasing the
2653–2669, 1995.
speed. If the time is exceeded, the robot sends an error log to [6] Fisher, M. L., Jörnsten, K. O. and Madsen, O. B. G. (1997), Vehicle
the routing system describing the problem found during the Routing with Time Windows: Two Optimization Algorithms,
route. Operations Research. Res, 45, 487-491.
[7] J. Z. Nanyang, W. Hsu, “Conflict-free container routing in mesh yard
layouts”, Robotics and Autonomous Systems, vol. 56, no. 5, pp. 451-
V. CONCLUSIONS 460, May. 2008.
[8] H. N. Psaraftis, “Dynamic vehicle routing problems”, Vehicle
Firstly, we described the most important works found in Routing: Methods and Studies. BL. Golden, AA. Assad,
literature that focus on routing systems for AGVs. Next, we editors.Vehicle routing: methods and studies. Amsterdam: North-
presented the approach used to develop our Routing system. Holland., pages 223–48, 1988.
[9] Ch.W. Kim and M. A. J. Tanchoco, “Conflict-free shortest-time
It is able to solve traffic jams and collisions. It also bidirectional AGV routing”, International Journal of Production
guarantees optimized routes before sending the final paths to Research, vol.29, n. 12, pp.2377–2391, 1991.
the robotic forklifts. It also verifies the progress of all tasks. [10] M. Gendreau, F. Guertin, J.Y. Potvin, E. Taillard, “Parallel tabu
search for real-time vehicle routing and dispatching,” Transportation
When a problem occurs, the routing system can change the Science, vol, 33, no. 4, pp. 381–390, 1999.
tasks priorities, routes, etc. to avoid new conflicts. In order [11] L. Qiu, W.J. Hsu, “A bi-directional path layout for conflict-free
to verify the algorithm’s performance, the robotic forklifts routing of AGVs,” International Journal of Production Research, vol.
9, no. 1, pp. 2177–2195, 2001.
were tested executing tasks that simulate the load and unload [12] S. Maza, P. Castagna, “Conflict-free AGV routing in bi-directional
of goods using the interface and the virtual environment in network”, in: Proc. 8th IEEE International Conf. on Emerging
the Player/Stage software. The use of the interface was very Technologies and Factory Automation, 2001.
interesting because it allowed us to control the process [13] S. Maza, P. Castagna, “Robust conflict-free routing of bi-directional
automated guided vehicles”, in: Proc. 9th IEEE International Conf.
simulation at a higher level. Therefore, this program can be on Emerging Technologies and Factory Automation, 2002.
used directly in industries at future. [14] R. H. Möhring, E. Köhler, E. Gawrilow and B. Stenzel, “Conflict-free
The A* algorithm was a simple way to control the robot Real-time AGV Routing”, Operations Research Proceedings, vol.
2004, Springer Berlin, 2004.
maneuvers, and the EKF was sufficient to avoid the position [15] M. Klimm, E. Gawrilow, R. H. Möhring and B. Stenzel, (Nov. 2008,
error propagation. Both algorithms were implemented in C. 19) “Conflict-free vehicle routing: load balancing and deadlock
As a result, we could solve many faults and refine the prevention,” Technical Report, 2007. Available:
<http://www.matheon.de/preprints/5137_preprint-static-routing.pdf>.
algorithms before embedding them in real robots. We also [16] E. Gawrilow, E. Köhler, R. H. Möhring and B. Stenzel, “Dynamic
plan some improvements on the algorithm, for instance, to Routing of Automated Guided Vehicles in Real-time”, Technische
refine the routing task planner and to take into account more Universit at Berlin, Book: Mathematics - Key Technology for the
future, pp. 165-177, 2008.
realistic parameters during the simulations (e.g.: velocity [17] S. Ravizza. Control of Automated Guided Vehicles (AGVs), diploma
changes while maneuvering, load and unload times, etc.). It thesis, ETH Zurich, 2009.
is important to emphasize that in the current version, the [18] I. F. A. VIS, “Survey of research in the design and control of
automated guided vehicle systems” in European Journal of
algorithm assigns the time for each path according to the Operational Research., vol. 170, Elsevier Science, Amsterdam , 2006,
distance, maneuvers and loading and unloading of pallets. pp. 677-709.
The robotic forklift speed during the path is determined by [19] J. Huang, U. S. Palekar, S. Kapoor, “A labeling algorithm for the
the local planner, which checks the distance to achieve the navigation of automated guided vehicles,” Journal of engineering for
industry, vol. 115, pp. 315–321, 1993.
goal and determines the speed required for the robotic [20] E. W. Dijkstra, “A Note on Two Problems in Conexion with Graphs”
forklift to arrive there in time. One limitation of the Numerische Math, 1959, vol. 1, pp. 269-271.
algorithm is that collisions are always solved by finding a [21] J. Desrosiers, F. Soumis, M. Desrochers and M. Sauvé. "Methods for
routing with time windows," European Journal of Operational
new route. A possible solution would be the application of Research, vol. 23, p. 236-245, 1986.
stoppages to one of the forklifts during a certain time [22] P. Serafini, "Dynamic programming and minimum risk paths",
interval, or the reduction of its speed. Both options will be European Journal of Operational Research, vol. 175, no. 1, pp. 224-
237, Nov. 2006.
investigated in future. At this moment our group is finishing [23] Nokia Corporation, (2010, Feb. 10). “Qt creator Whitepaper”, 2009
the design and construction of mini-mobile robots and a 1:5 [Online]. Available: http://qt.nokia.com/files/pdf/qt-creator-1.3-
scale warehouse to test experimentally our algorithms. whitepaper
[24] K.T. Vivaldini, J. P. M. Galdames, T. B. Pasqual, R. C. Araújo, R. M.
Sobral, M. Becker, and G. A. P. Caurin" Robotic Forklifts for
REFERENCES Intelligent Warehouses: Routing, Path Planning, and Auto-
[1] A. J. Broadbent, et al. “Free-ranging AGV and scheduling system,” localization", in IEEE – International Conference on Industrial
Automated Guided Vehicle Systems, 1987, pp.301–309. Technology, Viña del Mar – Valparaíso, Chile, Mar. 2010.
[2] L. D. Bodin, et al. “Routing and scheduling of vehicles and crews: [25] K.T. Vivaldini, J. P. M. Galdames, T. B. Pasqual, M. Becker, and G.
The state of the art,” Computers and Operations Research, vol. 10, no. A. P. Caurin, “Intelligent Warehouses: focus on the automatic routing
2, pp. 63–211, 1983. and path planning of robotic forklifts able to work autonomously,”
[3] Laporte, G., “The vehicle routing problem: An overview of exact and Mechatronics Systems: Intelligent Transportation Vehicles, 2010,
approximate algorithms,” European Journal of Operational Research, submitted for publication.
vol. 59, pp.345–358, 1992.