Mechatronics 45 (2017) 1–13

Contents lists available at ScienceDirect

Mechatronics
journal homepage: www.elsevier.com/locate/mechatronics

Cooperative cloud robotics architecture for the coordination of

multi-AGV systems in industrial warehousesR
Elena Cardarelli a, Valerio Digani b, Lorenzo Sabattini a,∗, Cristian Secchi a, Cesare Fantuzzi a
a
Department of Sciences and Methods for Engineering (DISMI), University of Modena and Reggio Emilia, Italy
b
Elettric80 s.p.a., via G. Marconi 23, 42030 Viano, Italy

a r t i c l e i n f o a b s t r a c t

Article history: In this paper we introduce a novel cloud robotics architecture that provides different functionalities to
Received 11 May 2016 support enhanced coordination of groups of Automated Guided Vehicles (AGVs) used for industrial logis-
Revised 15 February 2017
tics. In particular, we define a cooperative data fusion system that, gathering data from different sensing
Accepted 7 April 2017
sources, provides a constantly updated global live view of the industrial environment, for coordinating
the motion of the AGVs in an optimized manner. In fact, local sensing capabilities are complemented
Keywords: with global information, thus extending the field of view of each AGV. This knowledge extension allows
Logistics systems to support a cooperative and flexible global route assignment and local path planning in order to avoid
Multi-AGV systems congestion zones, obstacles reported in the global live view map and deal with unexpected obstacles in
Cloud robotics
the current path. The proposed methodology is validated in a real industrial environment, allowing an
Data fusion
AGV to safely perform an obstacle avoidance procedure.
Path-planning
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction gistic operations. AGV systems (Fig. 1) are employed for managing
the automatic transportation of raw materials and final products
Factory logistics is a crucial aspect for the production flow in among different locations of an industrial site and they are typ-
industrial plants: since the transportation of raw materials and ically employed to pick up a pallet of goods from the end of an
final products is a very frequent operation in a warehouse, any automated production line, and bring it to the warehouse, or from
bottleneck and inefficiency strongly affect the overall energy con- the warehouse to the shipment [3,4]. A complete survey relative to
sumption, impacting on the factory productivity and competitive- the principal technologies employed for the AGVs deployment in
ness on the market. In the last decades the goods production flow industrial plants is presented in [13]. The use of AGVs is becoming
in industrial environments has reached high levels of automation widespread in the last few decades: Amazon Robotics [5–7] is one
in all manufacturing processes, allowing to improve the safety for of the most popular solutions adopted in distribution centers for
workers, increase the efficiency of the production and reduce costs. handling the goods delivery. Because of the presence of huge and
However, factory logistics is still marginally integrated in auto- heavy loads to be managed [8], the complete automation in factory
mated manufacturing processes and it involves manual operations logistics is still not achievable within real production plants, there-
performed by human workers and manually driven forklifts: this fore, it is mandatory to have human workers and hand-driven fork-
arouses inefficiencies together with high risky working conditions lifts that cooperate with the AGV systems. In order to avoid dan-
for workers [1] and problems in product traceability. To cope with gerous working situations for human operators, AGVs are equipped
the lack of efficiency and flexibility in factory logistics, warehous- with certified laser scanners designed to guarantee the detection of
ing in modern industries can rely on Automated Guided Vehicles all unforeseen obstacles in the neighbourhood of each AGV. More-
(AGVs) [2], and integrated systems for the complete handling of lo- over, to enhance safety in the presence of unpredictable events, ad-
ditional constraints are applied to the motion of the AGVs, defining
a roadmap [9–12], that is a set of predetermined paths along which
AGVs are constrained to move (Fig. 2) in order to perform their
R
This paper was recommended for publication by Associate Editor Prof. Nariman tasks. As illustrated in [13], in most of modern automatic ware-
Sepehri. houses the use of a roadmap is a common strategy, that allows to
∗
Corresponding author.
reduce the complexity of the optimization problem for the coor-
E-mail addresses: [email protected] (E. Cardarelli),
[email protected] (V. Digani), [email protected] (L. Sabattini), dination of the AGV systems. The AGV fleets are typically handled
[email protected] (C. Secchi), [email protected] (C. Fantuzzi). by a centralized control system, usually referred to as Warehouse

http://dx.doi.org/10.1016/j.mechatronics.2017.04.005
0957-4158/© 2017 Elsevier Ltd. All rights reserved.
2 E. Cardarelli et al. / Mechatronics 45 (2017) 1–13

Fig. 1. AGVs fleet employed for the automated transportation of goods in a modern factory.

Fig. 2. Portion of a roadmap in a real industrial plant.

Management System (WMS), that is in charge of assigning the tasks data fusion, mission assignment, time synchronization, visual data
(usually referred to as missions) to the AGVs and coordinating their representation, global navigation and local path planning.
motion. The paper is organized as follows. In Section 2 we present an
The contribution of this paper concerns with the definition of overview of the system under consideration, highlighting the moti-
a centralized and cooperative architecture designed to enhance the vation of the principal technologies integrated to enhance the per-
coordination of AGVs in real industrial environments. Some prelim- formance. Subsequently, Section 3 introduces the main architecture
inary results have been presented in [14] and [15]. In this paper of the cloud system, whose implementation strategy is illustrated
we describe in detail the centralized cloud system responsible for in Section 4. The experimental validation of the system is reported
providing several services over the network of AGVs: multi-sensor
 E. Cardarelli et al. / Mechatronics 45 (2017) 1–13 3

in Section 5 and, finally, Section 6 contains some concluding re-

marks.

2. System overview

Despite the fact that the AGV systems are a very attractive so-
lution for increasing the level of automation in factory logistics,
their market share is still limited to few thousands vehicles sold
every year: some technological issues related to the plant instal-
lation costs, the flexibility of the coordination strategies and the
safety of the perception systems still need to be addressed to allow
a pervasive diffusion of AGVs in modern factories. The PAN-Robots
project focused on the development of novel sensing and control
technologies for AGV systems, to increase the degree of automa-
tion in factory logistics, allowing to improve the overall system
performance in terms of productivity, speed, accuracy, and flexi-
bility. The technological solutions proposed within PAN-Robots and
discussed in this paper concern with the development of a flexible
control system to handle the motion coordination of the AGVs, the
introduction of an advanced perception system to provide a safe
shared environment among the humans and the autonomous vehi-
cles and, finally, a centralized system, that is in charge of receiving
data from different sources, opportunely merging them, and mak-
ing them available for the AGV control system.

2.1. Global navigation and local path planning

Fig. 3. In the presence of an obstacle, the first AGV gets stuck, arousing traffic prob-
lems for the overall fleet of vehicles.
The definition of a roadmap is a very effective strategy for con-
trolling the global navigation of AGVs: it significantly simplifies the
coordination problem, especially when dealing with a high num- capable of classifying the detected objects, preventing the possi-
ber of AGVs. However, since AGVs are constrained to move on bility of differentiating the motion control of the AGV according
fixed paths, they are not allowed to deviate from the roadmap; to the type of obstacle in its sensing range. While static obsta-
therefore, in the presence of unforeseen obstacles it is necessary cles could be easily passed, without having any negative impact on
to find an alternative path on the roadmap, which is not always the overall safety of the system, conversely, if a human is crossing
possible as, for instance, in the frequent case of mono-directional the sensing range of an AGV, the only safe procedure is to avoid
roads. In this situation, the AGV is forced to stop, until the obsta- any movement. However, these high level decision strategies could
cle has been removed. This lack of flexibility strongly impacts on not be implemented because laser scanners do not allow to dis-
productivity and huge traffic problems can be generated, especially tinguish between humans and other kind of obstacles. Moreover,
in the presence of multiple AGVs (Fig. 3). To cope with this prob- since obstacle detection is limited to the neighborhood of each
lem, a functionality (referred to as Local Path Planning) for comput- AGV, global information on the surrounding environment is not
ing local deviations from the roadmap is introduced in [16]: con- available, hence, to avoid collision with unexpected moving objects
sidering the dynamic and kinematic constraints of the AGVs, the in critical zones as, for instance, intersections and blind spots, the
proposed algorithm is based on the computation of polar spline AGV is forced to reduce its speed.
curves, lane-change maneuver curves and line segments to obtain As a motivating example, consider the scenario in Fig. 4a: in
new obstacle-free trajectories that leave the original roadmap pro- the presence of multiple objects, as well as pedestrians, an AGV is
viding collision avoidance. moving on the path of roadmap represented by the orange dashed
line. The on-board perception system allows to identify only a
2.2. Mission manager limited portion of the obstacles in the neighborhood of the AGV
(green area in Fig. 4a). Therefore, based only on local sensing, it is
The algorithm introduced within PAN-Robots to manage task possible to compute an obstacle-free deviation from the roadmap
assignments and to subsequently plan the path to be traveled for (blue line in Fig. 4b). However, as represented in Fig. 4c, exploiting
mission completion, is based on the minimization of the overall the local sensing capabilities could not be sufficient to guarantee
completion time, considering different factors that influence it, e.g. collision avoidance: following the obtained maneuver, the AGV is
AGV that is closest to the required operation point, congestion sit- in conflict with a previously occluded pedestrian, not directly visi-
uations in the warehouse, required battery exchanges. ble by the on-board perception system during the computation of
The proposed methodology, described in details in [17], consists the local deviation (Fig. 4a).
in exploiting the Hungarian Algorithm, where activities are repre- To cope with these issues, an advanced perception system is
sented by missions to be accomplished, and agents are represented proposed in order to support a reliable obstacle detection, track-
by AGVs. ing and classification, providing a safe shared environment among
the humans and the vehicles.
2.3. Perception system The sensing system proposed within PAN-Robots is composed
by:
Because of the presence of human workers and manually driven
forklifts in the warehouse, certified laser scanners are installed on- • A semi-autonomous plant exploration system, in the form of an
board each AGV, providing a reliable and robust obstacle detection advanced AGV equipped with multiple laser scanners (Fig. 5a),
in the vehicle surrounding area. However, these sensors are not capable of computing automatically a 3D map of the ware-
4 E. Cardarelli et al. / Mechatronics 45 (2017) 1–13

(a) Multiple-objects scenario with an (b) Computation of an obstacle-free tra-

AGV ready to follow the predefined road jectory (blue line) from the predetermined
represented by the orange dashed line. roadmap (orange dashed line) relying only
Based on local sensing, obstacle recogni- on local sensing capabilities.
tion is limited to the vehicle surrounding
area (green region) and some obstacles,
such as the pedestrian and the small box,
are not directly detected because they are
occluded by the big box.

(c) Obstacle collision in the computation (d) Global information from the cloud sys-
of local deviation exploiting local sensing. tem.
Fig. 4. Motivating example for the development of advanced perception systems.

(a) Semi-autonomous (b) Laser based infras- (c) On-board perception

plant exploration tructure perception
Fig. 5. Advanced perception system.

house, representing the geometric layout of the plant itself, in- stacle detection, tracking and classification [19] with a position
cluding the position of all points of interest and structural el- and velocity tracking accuracy of, respectively, 0.098 m at 1σ
ements placed inside it [18]. The system provides a mapping and 0.22 m/s at 1σ .
accuracy of ±5 cm at 1σ . • An on-board perception system, composed by a safety laser scan-
• A laser based infrastructure perception system, used to monitor ner, supplemented by another laser scanner and by an omni-
intersections and blind spots (Fig. 5b), providing runtime ob- directional stereo vision subsystem consisting of two omnidi-
 E. Cardarelli et al. / Mechatronics 45 (2017) 1–13 5

rectional lenses and two cameras mounted on the top of the with in an automated manner, exploiting sensing devices, in sev-
AGV (Fig. 5c). The system allows to detect, classify and track eral cases the human intervention is mandatory. Hence, it is nec-
objects [20] in the 360° area around the AGV: 11 m in front of essary for the system to provide human operators with an efficient
the AGV, 8 m in the back and 5 m on the left/right side. The supervision interface, to let them effectively assess the behavior of
detection, tracking and classification performance were evalu- the overall system. The Human Machine Interface (HMI) proposed
ated in different scenarios [21]: the provided detection rate is within PAN-Robots [22] provides the operator with a constantly
92.8%, with a correct tracking rate of 95%. The measured av- updated three-dimensional view of the overall system, giving the
erage localization error is 3% at 1σ and the velocity error is workers the possibility of being aware, in an intuitive manner, of
± 0.38 m/s at 1σ . The running time is about 95 ms. the behavior of the multi-vehicle system. The operator also has the
possibility of modifying this behavior, defining custom missions to
be executed.
2.4. Global Live View
3. Control Center
As shown in Fig. 4, system safety could not be guaranteed when
the collision avoidance strategy of a specific AGV only relies on
All the functionalities developed within PAN-Robots to sup-
the data perceived in the single vehicle surrounding area: there-
port intelligent AGV navigation are combined in the Control Center
fore, it is necessary to support data sharing among AGVs in or-
(Fig. 6), that is responsible for:
der to increase the safety level of the system. A possible solu-
tion could be based on the broadcasting towards the fleet of all • Improving safety and efficiency in the computation of local de-
data locally perceived by each AGV: integrating data coming from viation maneuvers for avoiding obstacles (Local Path Planning);
the infrastructural sensors and from the other AGVs, each vehicle • Optimizing the motion coordination of the AGVs for accom-
is constantly updated about the status of the whole plant. Since, plishing their tasks (Autonomous Global Navigation);
for this solution, a large quantity of data need to be simultane- • Offering to the PAN-Robots AGV an updated map of the indus-
ously transmitted over the network, it is necessary to reserve a trial environment (Global Live View);
huge amount of communication bandwidth in order to cope with • Providing an intuitive Human Machine Interface (HMI), that al-
the high transmission overhead and provide a reliable communi- lows human operators to interact with the system and receive
cation among AGVs. Moreover, since obstacle information is pro- constantly updated information regarding the fleet of AGVs and
vided by several sources, each AGV needs to be equipped with high the status of the industrial environment;
performance processing units, in charge of calculating the collision • supervising the AGV fleet by managing their missions and dif-
avoidance strategy combining and processing all data received and ferent types of operation points (Mission Manager).
resolving problems as data redundancy, inconsistency, ambiguity,
As represented in Fig. 6, the architecture of the Control Center
noise and incompleteness: equipping an industrial fleet of AGVs
is divided in a data processing layer, where the processing of re-
with high performance computation systems may involve a sig-
ceived data takes place, and an application layer that provides ac-
nificant cost for the factory. While the communication overhead
cess to external entities, like a human operator or the Warehouse
within broadcasting based solutions could be reduced limiting the
Management System.
data exchange among close AGVs, the costs related to the AGVs
In the data processing layer, the Control Center manages the
equipment and setup could be reduced introducing, within the sys-
layered map provided by the plant exploration system and adds,
tem, a separate module in charge of collecting all data perceived
through the Global Live View, a further layer for the representa-
by the sensors and combining them in a unique and complete rep-
tion of the dynamic objects perceived by the on-board and infras-
resentation to be shared to the AGVs. This centralized system, re-
tructure sensing systems. In the application layer the Mission Man-
ferred to as Global Live View [14,15], enhances the quality of the
ager module is in charge of assigning missions from the Ware-
information providing to the AGV control system a global updated
house Management System (or manually defined using the HMI
map representing the static entities (the 3D map of the plant, the
connected via Ethernet) to each AGV. Based on the mission as-
route map), the dynamic entities (the current position and veloc-
signment, the motion of the AGVs on the roadmap is coordinated
ity of the AGVs, the position and velocity of currently identified
by the Global Navigation. All relevant information is distributed to
objects), the congestion zones and the status of the monitored in-
the AGVs via wireless communication in a local network. Beside
tersections.
managing the interaction of the components, the Control Center is
The Global Live View extends the local sensing capabilities, in-
moreover responsible for synchronizing the AGVs and infrastruc-
tegrating data coming from all perception systems, namely the on-
ture perception systems, distributing via wireless a global time ref-
board and the infrastructure ones (Fig. 4d), allowing to improve
erence through the Time Server service.
the overall safety of the system. The choice of a centralized system
The PAN-Robots system can be considered as cloud robotics
is preferable for industrial environments where preventing danger-
architecture according to the definition provided by the National
ous situations is a priority with respect to the lack of flexibility:
Institute of Standards and Technology (NIST), that defines the
for this reason, in the presence of an obstacle, when the commu-
cloud computing as “a model for enabling ubiquitous, convenient,
nication with the Global Live View is poor or broken, obstacle de-
on-demand network access to a shared pool of configurable resources
tection is demanded to the local sensing of each AGV, forcing the
(e.g., servers, storage, networks, applications, and services) that can be
vehicle to stop and wait until the obstacle has been removed.
rapidly provisioned and released with minimal management effort or
service provider interaction” [23]. As the DAvinCi server of the cloud
2.5. Human Machine Interface computing framework presented in [24], the Control Center acts
as a service provider that collects data from heterogeneous net-
Since industrial plants are highly dynamic environments, shared worked agents represented by AGVs, users and perception systems
among AGVs and human workers, they define a very complex sys- and provides them all the functionalities presented as real time
tem, difficult to be controlled in a fully automated manner: distur- services. Therefore, the Control Center runs several services imple-
bances on the nominal behavior of the system may appear, such as mented as parallel threads communicating via socket over the pri-
the presence of obstacles on the AGVs’ path, or necessity to change vate cloud; the use of a private network is motivated by security
missions for the AGVs. While the presence of obstacles can be dealt constraints: since the environment is shared among vehicles and
6 E. Cardarelli et al. / Mechatronics 45 (2017) 1–13

Fig. 6. The cooperative architecture of the Control Center as a centralized cloud service.

humans, disrupted functionalities and damaged facilities caused by and by the infrastructure perception system and sends to the HMI
undesired remote attacks could generate dangerous situations for the list of classified entities to be graphically represented, namely
workers. In the designed architecture machine-to-machine com- infrastructural elements, AGVs, static and dynamic objects.
munication (M2M in [25]) among the AGVs is not implemented: The coordination of the AGVs (Fig. 8) is composed of two main
for safety reasons, in case of failures or unavailable access it is modules: Mission Manager and Global Navigation.
preferable adopting a conservative approach, where each AGV re- The list of operational tasks to be performed by the fleet
lies on its local sensing and stops until an object has been re- of AGVs are collected and stored into the Warehouse Manage-
moved, rather than handling high latencies generated by routing ment System (WMS), that is a centralized and constantly updated
protocols [26] necessary to maintain the reachability of each node database, connected to the production system. The Mission Man-
over a M2M communication network. ager is then in charge of querying the WMS to receive missions,
Providing the Global Live View service, the Control Center al- and assign them to specific AGVs. Mission assignment is performed
lows the AGVs to share and extend their knowledge about the en- exploiting the Hungarian Algorithm that, as is well known, repre-
vironment, integrating local sensing data with information coming sents the optimal algorithm for solving the assignment problem.
from the other perception systems (Fig. 4d). Therefore, exploiting Specifically, the Hungarian Algorithm is utilized for assigning a cer-
completely the perception capabilities of the whole network, it is tain number of activities (i.e. the missions) to a certain number of
possible to solve cooperatively some complex tasks, as local path agents (i.e. the AGVs). The proposed methodology, described in de-
planning computation and global navigation. tails in [17], performs assignment minimizing the overall cost, that
According to the definition in [27] a cloud robot and automa- do not depend only on the distance to be traveled, but also on the
tion system is “Any robot or automation system that relies on ei- current traffic condition.
ther data or code from a network to support its operation, i.e., where Based on the result of the mission assignment, the Global Nav-
not all sensing, computation, and memory is integrated into a single igation module is in charge of defining the path to be traveled by
standalone system”; although the Control Center is the core of the each AGV along the roadmap. Specifically, global navigation is the
cloud computing architecture, not all computation, memory and problem of assigning a path along the roadmap to each AGV, in
programming is concentrated in it. For instance, obstacle detection such a way that missions are completed while minimizing the to-
for the reconstruction of global live view is performed indepen- tal traveling time and avoiding conflicts. The problem consists then
dently by the AGVs, that are focused on sensing their neighbor- of planning a path for a fleet of AGVs in a 2D static environment
hood with a low consumption of resources, while data fusion and to avoid conflicts and deadlocks. Each AGV starts its path in an ini-
computationally intensive algorithms are delegated to the Control tial position, and has to reach its own final position. The proposed
Center. The capability of offloading computationally expensive al- methodology, described in details in [11], solves the path planning
gorithm to the cloud allows to save costs employing AGVs with problem on two levels:
low computation power and memory requirements: this is one of
the benefits of the cloud computing, as highlighted in [25]. 1. In the high level, or Topology level, the environment is parti-
The UML sequence diagrams in Figs. 9, 8 and 7, represent the tioned into sectors, that are bounded areas of limited dimen-
interaction among the principal threads running on the Control sion. The planning algorithm then searches for the best path to
Center: Global Live View, Local Path Planning, Autonomous Global the final goal (actually to the final sector where the real goal
Navigation, Mission Manager and HMI. is) from the current sector. A receding horizon approach is uti-
In order to provide human workers with an intuitive and ef- lized, together with a graph search algorithm, to allow path re-
fective HMI (Fig. 7), the Global Live View fuses runtime the infor- planning if necessary. The path on this layer is composed of a
mation about the obstacles detected by the AGV on-board system sequence of sectors the AGV has to cross.
 E. Cardarelli et al. / Mechatronics 45 (2017) 1–13 7

Fig. 7. UML diagram for human intuitive data representation.

Fig. 8. UML diagram for the motion coordination of the AGVs.

Fig. 9. UML diagram for the computation of local deviations.

2. In the low level, or Roadmap level, planning is performed 4. Data fusion

within each sector. The path is computed for each AGV on the
roadmap, and coordination is handled to avoid conflicts among Multisensor data fusion is performed within the Global Live
the AGVs, exploiting a priority scheme. View in order to merge data acquired by the different sensors,
maintaining consistency with respect to the real world: each vir-
tual object represented in the map must have a correspondence to
a real object of the world. Multisensor data fusion has been deeply
In the current state-of-the-art AGV systems, no local deviation
investigated in the recent years: a complete review of contempo-
from the predefined roadmap is allowed. When an AGV get stuck
rary data fusion methodologies, as well as the most recent devel-
on a segment, it has to wait until the obstacle has been removed.
opments and emerging trends is presented in [28]. Different crite-
This is not efficient and very time consuming for the whole system.
ria can be used for the classification of the data fusion techniques,
A methodology was then proposed for solving this issue, comput-
as discussed in [29]. According to the level of abstraction we can
ing local deviations (Fig. 9) when the on-board perception system
distinguish these classes:
detects an obstacle on the AGV’s path. Based on the obstacles’ po-
sitions, and on the characteristics of the roadmap, opportune de-
viations are computed, utilizing the algorithm detailed in [16], to • Low level fusion (measurements level): this level deals di-
let the AGV overtake the obstacle. In particular, when an obsta- rectly with raw data to improve the accuracy of the individual
cle has been detected, the AGV sends to the Local Path Planning a sources.
request for computing a local deviation from the roadmap. The re- • Medium level fusion (features level): it is based on the process-
quest is then forwarded to the Global Live View, that computes the ing of features or characteristics (dimension, shape, position).
list of obstacles in the neighborhood of the AGV and sends it back • High level fusion (decisions level): also known as decision fu-
to the Local Path Planning. Then, this module computes the curve sion, this level addresses symbolic representation, such as ob-
to implement the local deviation in a safe manner, evaluating its ject classes. This level is also known as the feature or charac-
feasibility and transmitting this information toward the AGV. teristic level.
8 E. Cardarelli et al. / Mechatronics 45 (2017) 1–13

• Multiple level fusion: based on the processing of data provided Algorithm 1 Geometric level data fusion.
at different levels of abstraction.
Require: obj_list_from_infrastructure, obj_list_from_onboard
Considering the point where data fusion is performed it is pos- 1: procedure Data alignment

sible to distinguish three types of architectures: centralized, dis- 2: for all infra_obj ∈ obj_list_from_infrastructure do
tributed and hierarchical. A centralized architecture is based on 3: Convert_to_Global_reference(infra_obj)
the presence of a single node responsible for making decisions 4: for all on_board_obj ∈ obj_list_from_on_board do
according to the data collected from the input sources. The pos- 5: Convert_to_Global_reference(on_board_obj)
sibility of communication bottlenecks and the large consumption 6: procedure Data association
of bandwidth for transmitting raw data over the network are the 7: for all infra_obj ∈ obj_list_from_infrastructure do
principal drawbacks of this solution. In a distributed architecture 8: for all on_board_obj ∈ obj_list_from_on_board do
each single source node processes independently raw data and 9: Evaluate_overlap(infra_obj, on_board_obj)
provides to the multisensor fusion a local estimation of the object 10: return obj_fused_list
status. In hierarchical architectures data fusion is performed at dif- 11: procedure Static and dynamic discrimination
ferent levels, combining decentralized and distributed nodes. Dif- 12: for all fused_obj ∈ obj_fused_list do
ferent approaches have been considered to implement multisensor 13: if velocity(fused_obj ) < th then
data fusion within the Global Live View [14], such as occupancy 14: fused_obj is static
grids [30,31], feature-based approaches [32] and octrees [33]. 15: else
Compared with feature-based approaches, occupancy grids al- 16: fused_obj is dynamic
low to reach more flexibility and robustness in the fusion of noisy
information, allowing to integrate heterogeneous input data in the
same framework according to the inherent uncertainty of each in-
according to a global positional and temporal reference system
put sensor. On the other hand, the mapping algorithms commonly
(lines 3 and 5 in Algorithm 1). Then, processing the position and
employed for updating the cells of the occupancy grids, such as
orientation of the bounding boxes related to the area covered by
fast inverse models [31,34] or forward models[35] involve costly
the detected objects, it is possible to compute their occupational
operations and large memory requirements, especially during their
overlapping (line 9) and discriminate among bounding boxes rep-
initialization and updating phases, when the map is extended and
resenting different views of the same object or separated elements.
the mapped environments are not known beforehand. Octrees al-
This allows to obtain a preliminary list of fused objects (line 10),
low to cope with these limitations but despite the accuracy, flexi-
on which the information about the estimated velocities and di-
bility and robustness provided, all these approaches consist in the
rections are integrated in order to discriminate among static and
fusion of low level data (images, 3D point clouds, laser raw data)
dynamic obstacles (see procedure in line 11). The information rep-
and for the application we are considering we assume to pro-
resenting the fused obstacles detected runtime is then fused with
cess, for each obstacle candidate, geometric/tracking features (ID,
the static 3D map processed by the semi-autonomous plant explo-
age, position, orientation, velocity and size) and decision level fea-
ration system, modeling a grid map on which free space, occupied
tures (class and classification quality) in order to optimize the data
and unknown regions are represented. This allows to support both
transmission time and reduce the network overhead. Hence, the
local path planning and autonomous global navigation capabilities.
Global Live View is implemented as a hierarchical data fusion ar-
chitecture that deals separately with geometric and decision level
data fusion. 4.2. Decision fusion

From a high level point of view the implementation of the

4.1. Geometric fusion
Global Live View is considered as a classifier combination prob-
lem: the details of the developed procedure are illustrated in
Since object tracking is performed both by the on-board multi-
Algorithm 2. Processing motion, directional and dimensional fea-
sensor systems of the AGVs and by the laser infrastructure-based
perception systems we considered to address the geometric fu-
sion problem as a target tracking process implementing distributed Algorithm 2 Decision data fusion.
track fusion methodologies as maximum-likelihood [36] and min- Require: obj_fused_list
imum mean square error (MMSE) solutions. Since the distribution 1: procedure Pre-Classification
of tracking task among multiple sources involves the possible in- 2: for all fused_obj ∈ obj_fused_list do
troduction of correlated errors among the tracks [37], additional 3: if fused_obj is static then
processing is required, as the estimation of the cross covariances 4: Evaluate(size(fused_obj))
between track estimates. This calculation is computationally ex- 5: else
pensive, hence, to address this problem it is possible to employ 6: Evaluate(size(fused_obj), velocity(fused_obj))
methodologies based on direct track-to-track fusion [38] or, alter- 7: return pre_classified_list
natively, treating the decorrelation of the state estimates [39]. In 8: procedure Classification
both cases, additional information as covariances and correspond- 9: for all pre_classified_obj ∈ pre_classified_list do
ing measurement matrices must be sent to the global level: this 10: Calculate_Class_Votes_Average(pre_classified_obj)
is not feasible for the application we are considering because of
11: return classified_list
the limited transmission capabilities of the wireless network in-
frastructure.
Hence, we introduce a heuristic, based on the evaluation of tures (lines 3, 4, 5 and 6) a preliminary classification value is calcu-
the occupational overlapping among the different lists of objects lated for each element of the fused list obtained at geometric level
detected by the infrastructure and on-board perception systems. (Algorithm 1). Then, the static 3D map of the environment, the
The principal steps of the procedure are reported in Algorithm 1. on-board sensor systems and the infrastructure perception systems
Firstly, the input lists of objects (i.e. obj_list_from_infrastructure and are considered as a set of classifiers that, given an unknown ob-
obj_list_from_onboard) are geometrically and temporally aligned, ject to classify, provide an output score, that is a confidence value
 E. Cardarelli et al. / Mechatronics 45 (2017) 1–13 9

(a) Sensing area covered by the infrastruc- (b) Field of view of the on-board extero-
tural laser scanner (in green) ceptive sensors (in red)
Fig. 10. Sensing area of the advanced perception systems.

Fig. 11. Evaluation scenario observed from the on-board sensing system.

for each possible class of the system (human, manual forklift, AGV, Table 1
Single transmission scenario.
other dynamic object, static object) to be the correct class for the
unknown input pattern. The classification results provided by the Message type Description Value
sensors are combined with the rough class values obtained at geo- Pallet position
metric level, exploiting, at measurement level [40], the average-rule Update rate: 50 Hz
aggregation scheme: for each fused object, the average of the scores Packet loss: 0%
assigned to each class is computed, and the most voted class is Message correctness: 100%
Received messages: 4278
considered the correct one for the object. Despite its simplicity, the
Sent messages: 4278
average-rule aggregation scheme competes with the most sophisti- Operation point
cated combination methods, as highlighted in [41]. Update rate: 50 Hz
Packet loss: 0%
5. Results Message correctness: 100%
Received messages: 4314
Sent messages: 4314
Two principal aspects have been considered for the validation Route
of the system: first, demonstrating that the presented architecture Update rate: 50 Hz
is based on a reliable, robust and efficient communication network Packet loss: 0%
and, second, highlighting with an experimental scenario the ad- Message correctness: 100%
Received messages: 4156
vantages in terms of safety and flexibility introduced by the cloud
Sent messages: 4156
based methodology.

5.1. Performance evaluation

• A worst case for typical operational scenario where, with an
For evaluating the network capabilities we tested singularly and update rate of 10 Hz, we transmitted, every 60 s, 600 data pack-
simultaneously message exchanging among the principal modules ets for each message type and then we measured the network
of the system, measuring the packet loss rate and the data format capabilities.
correctness. Since in a typical operational scenario the message ex- • A stress testing where, doubling the typical update rate, we
changing frequency is less than 10Hz, for testing the single trans- transmitted at 20 Hz, every 60 s, 1200 messages of each mes-
mission scenario we have considered two cases: sage type and then we measured the network capabilities.
10 E. Cardarelli et al. / Mechatronics 45 (2017) 1–13

Since for both cases we obtained 0% of packet loss and 100% • Route: it represents the map of the segments that the AGVs
data correctness, we increased to 50 Hz the data transmission have to follow during the path.
rate, obtaining the results summarized in Table 1. Particularly,
Table 1 shows the data measured for the most computationally ex-
pensive and frequently exchanged message types: Concerning the test with simultaneous message exchanging, we
considered a sending rate of 20 Hz measuring the network per-
• Pallet position: it defines the geometric position of the pallet to formance after 30 min: about 36,0 0 0 packets were simultaneously
pick/place. transmitted with 100% correctness and 0% of packet loss. As for
• Operation point: it represents the position that the AGV has to the single transmission scenario, to further stress the system the
reach during the picking/placing operation. It does not corre- update rate has been increased to 50 Hz: the results obtained for
spond to the geometric position of the pallet, but it is the ge- the most frequently exchanged and computational costly messages
ometric position of the logical point, that is the point in which are summarized in Table 2. Finally, the sending rate has been in-
the AGV can stop to pick/place the load. creased to 120 Hz: in this case the measured packet loss is about

(a) Infrastructure sensing system

(b) On-board sensing system

Fig. 12. Bounding boxes of the objects acquired.
 E. Cardarelli et al. / Mechatronics 45 (2017) 1–13 11

1%; this is a limit of the software for data transmission and it is 5.2. Experimental validation
not related to the network congestion.
Moreover, we tested the transmission latency relative to obsta- The experimental validation of the proposed methodology was
cle detection, measuring, in presence of an obstacle, the difference carried on in a real industrial warehouse, where measurements
between the timestamp of the message sent by the on-board sys- were acquired by:
tem to Global Live View and the timestamp of the message relative
to the updated obstacles list sent by the Global Live View to the • An infrastructural laser scanner, mounted on a corner of the
AGV: the measured latency, after 50 tests, is averagely 43ms with warehouse (whose field of view is represented in green in
a standard deviation of 33ms. Fig. 10a);
All the obtained results demonstrate that the communication • An on-board sensing system, namely laser scanners and cam-
network is suitable for a real operational scenario, where the re- eras, installed on an AGV (whose sensing area is highlighted in
quested update rate is 10 Hz. red in Fig. 10b).

(a) Aligned input

(b) Data fusion output

Fig. 13. Bounding boxes of the aligned objects and output of the data fusion.
12 E. Cardarelli et al. / Mechatronics 45 (2017) 1–13

Table 2 Acknowledgment
Simultaneous transmission scenario.

Message type Description Value This paper is written within PAN-Robots project. The re-
Pallet position search leading to these results has received funding from the
Update rate: 50 Hz European Union Seventh Framework Programme (FP7/2007-2013)
Packet loss: 0% under grant agreement n. 314193.
Message correctness: 100%
Received messages: 7537
Sent messages: 7537 References
Operation point
Update rate: 50 Hz
[1] Eurostat. http://ec.europa.eu/eurostat;
Packet loss: 0%
[2] Sabattini L, Digani V, Secchi C, Cotena G, Ronzoni D, Foppoli M, et al. Tech-
Message correctness: 100% nological roadmap to boost the introduction of AGVS in industrial applications
Received messages: 7651 IEEE international conference on intelligent computer communication and pro-
Sent messages: 7651 cessing (ICCP); 2013.
Route [3] Mahadevan B, Narendran TT. Design of an automated guided vehicle-based
Update rate: 50 Hz material handling system for a flexible manufacturing system. Int J Prod Res
Packet loss: 0% 1990;28(9):1611–22.
Message correctness: 100% [4] Berman S, Edan Y. Decentralized autonomous AGV system for material han-
Received messages: 7322 dling. Int J Prod Res 2002;40(15):3995–4006.
Sent messages: 7322 [5] Wurman PR, D’Andrea R, Mountz M. Coordinating hundreds of cooperative, au-
tonomous vehicles in warehouses. AI Mag 2008;29(1):9.
[6] D’Andrea R. Guest editorial: a revolution in the warehouse: a retrospective
on kiva systems and the grand challenges ahead. IEEE Trans Autom Sci Eng
2012;9(4):638–9.
The sensing systems were used to acquired information about [7] Guizzo E. Three engineers, hundreds of robots, one warehouse. Spectr IEEE
static and dynamic entities in the environment. In particular, 2008;45(7):26–34.
[8] Oleari F, Magnani M, Ronzoni D, Sabattini L. Industrial AGVs: toward a perva-
Fig. 11 shows an evaluation scenario observed from the on-board sive diffusion in modern factory warehouses. In: Proceedings of the IEEE in-
perception system. The bounding boxes of the objects detected by ternational conference on intelligent computer communication and processing
the infrastructure and by the on-board sensing systems are shown, (ICCP). Cluj-Napoca, Romania; 2014. p. 233–8.
[9] Digani V, Sabattini L, Secchi C, Fantuzzi C. Hierarchical traffic control for
respectively, in Fig. 12a and b. The centralized data fusion cloud
partially decentralized coordination of multi agv systems in industrial envi-
aligns these data according to a global reference and temporal ronments. IEEE international conference on robotics and automation (ICRA);
frame (Fig. 13a) in order to allow the processing of their geometri- 2014a.
[10] Digani V, Sabattini L, Secchi C, Fantuzzi C. An automatic approach for the gen-
cal overlaps (Fig. 13b) and reducing noisy contribution.
eration of the roadmap for multi-agv systems in an industrial environment. In:
Concerning the performance evaluation, the measured update IEEE International Conference on Intelligent Robots and Systems (IROS); 2014b.
rate of the Global Live View is approximately 60 Hz, and the over- [11] Digani V, Sabattini L, Secchi C, Fantuzzi C. Ensemble coordination approach in
all processing time is 15 ms. The overall classification rate is equal multi-agv systems applied to industrial warehouses. IEEE Trans Autom Sci Eng
2015;12(3):922–34.
to 93.2%. [12] Beinschob P, Meyer M, Reinke C, Digani V, Secchi C, Sabattini L. Semi-auto-
The results relative to the validation of the local path planning mated map creation for fast deployment of AGV fleets in modern logistics.
strategies are shown in the accompanying video. It is important to Robot Auton Syst 2017;87:281–95.
[13] Vis IF. Survey of research in the design and control of automated guided vehi-
note that without advanced sensing and data fusion technologies, cle systems. Eur J Oper Res 2006b;170(3):677–709. doi:10.1016/j.ejor.2004.09.
in the presence of an obstacle the AGV is forced to stop until a hu- 020.
man operator manually removes it to allow the AGV to move on [14] Cardarelli E, Sabattini L, Secchi C, Fantuzzi C. Multisensor data fusion for ob-
stacle detection in automated factory logistics. IEEE international conference
its predetermined route. With the proposed technological solution, on intelligent computer communication and processing (ICCP). IEEE; 2014.
the AGV is able to autonomously and safely compute local devia- [15] Cardarelli E, Sabattini L, Secchi C, Fantuzzi C. Cloud robotics paradigm for en-
tions to pass the obstacle. hanced navigation of autonomous vehicles in real world industrial applica-
tions. In: Proceedings of the IEEE/RSJ international conference on intelligent
robots and systems (IROS). Hamburg, Germany; 2015.
[16] Digani V, Caramaschi F, Sabattini L, Secchi C, Fantuzzi C. Obstacle avoidance for
industrial AGVs. In: Proceedings of the IEEE international conference on intel-
6. Conclusion
ligent computer communication and processing (ICCP). Cluj-Napoca, Romania;
2014c. p. 227–32.
In this paper we introduced a cloud robotics architecture that [17] Sabattini L, Digani V, Lucchi M, Secchi C, Fantuzzi C. Mission assignment for
gathers information from different sensing sources. A global live multi-vehicle systems in industrial environments. IFAC symposium on robot
control (SYROCO). Salvador, Brazil; 2015.
view of the environment is then defined, that contains informa- [18] Beinschob P, Reinke C. Advances in 3d data acquisition, mapping and local-
tion regarding all the entities in the industrial environment. This ization in modern large-scale warehouses. In: Intelligent computer commu-
system is then used for enhancing the local sensing capabilities of nication and processing (ICCP), 2014 IEEE international conference on; 2014.
p. 265–71.
the AGVs, thus increasing the efficiency and the flexibility of the [19] Boehning M. Improving safety and efficiency of agvs at warehouse black spots.
AGV motion coordination. In: Intelligent computer communication and processing (ICCP), 2014 IEEE in-
Multisensor data fusion is achieved, to define the global live ternational conference on; 2014. p. 245–9.
[20] Drulea M, Szakats I, Vatavu A, Nedevschi S. Omnidirectional stereo vision using
view, combining sensing data after local processing, that provides fisheye lenses. In: Intelligent computer communication and processing (ICCP),
geometric and decision level data. 2014 IEEE international conference on; 2014. p. 251–8.
The proposed architecture was validated in a real industrial en- [21] Vatavu A, Costea D, Nedevschi S. Modeling and tracking of dynamic obsta-
cles for logistic plants using omnidirectional stereo vision. In: Proceedings of
vironment, where multiple sensors were deployed. In particular,
the IEEE/RSJ international conference on intelligent robots and systems (IROS).
the global live view was utilized for letting an AGV safely perform Hamburg, Germany; 2015.
an obstacle avoidance procedure. [22] Cardarelli E, Sabattini L, Digani V, Secchi C, Fantuzzi C. Interacting with a multi
AGV system. In: Proceedings of the IEEE international conference on intelligent
While the proposed architecture considered only warehouse en-
computer communication and processing (ICCP). Cluj-Napoca, Romania; 2015.
vironments, the concept can be extended to more generic intelli- [23] Mell PM, Grance T. Sp 800-145. the nist definition of cloud computing. Tech.
gent factory applications. Along these lines, future work will aim Rep.. Gaithersburg, MD, United States; 2011.
at including, inside the global live view, information regarding the [24] Arumugam R, Enti VR, Bingbing L, Xiaojun W, Baskaran K, Kong FF, et al.
Davinci: a cloud computing framework for service robots. In: Robotics and
status of the production lines and machines, providing support for automation (ICRA), 2010 IEEE international conference on; 2010. p. 3084–9.
optimization of the overall factory workflow. doi:10.1109/ROBOT.2010.5509469.
 E. Cardarelli et al. / Mechatronics 45 (2017) 1–13 13

Lorenzo Sabattini is an Assistant Professor at the De-

[25] Hu G, Tay WP, Wen Y. Cloud robotics: architecture, challenges and applica-
partment of Sciences and Methods for Engineering, Uni-
tions. IEEE Netw 2012;26(3):21–8.
versity of Modena and Reggio Emilia, Italy, since 2013.
[26] Shah D. Gossip algorithms. Found Trendsë Netw 2008;3(1):1–125. doi:10.1561/
He received his B.Sc. and M.Sc. in Mechatronic Engineer-
130 0 0 0 0 014.
ing from the University of Modena and Reggio Emilia
[27] Kehoe B, Patil S, Abbeel P, Goldberg K. A survey of research on cloud robotics
(Italy) in 2005 and 2007 respectively, and his Ph.D. in
and automation. IEEE Trans Autom Sci Eng 2015;12(2):398–409. doi:10.1109/
Control Systems and Operational Research from the Uni-
TASE.2014.2376492.
versity of Bologna (Italy) in 2012. In 2010 he has been
[28] Khalegh B, Khamis A, Karray FO, Razavi S. Multisensor data fusion: a review of
a Visiting Researcher at the University of Maryland, Col-
the state-of-the-art. Inf Fus 2013;14(1):28–44.
lege Park, MD (USA). His main research interests include
[29] Castanedo F. A review of data fusion techniques. Sci World J 2013;2013.
multi-robot systems, decentralized estimation and control
[30] Elfes A. Using occupancy grids for mobile robot perception and navigation.
control, and mobile robotics. He is one of the founding
Comput (Long Beach Calif) 1989;22(6):46–57.
co-chairs of the IEEE RAS Technical Committee on Multi-
[31] Moravec H. Sensor fusion in certainty grids for mobile robots. AI Mag
Robot Systems: he has served as the corresponding co-chair since its foundation, in
1988;9(2):61–74.
2014. He has been serving as Associate Editor for the IEEE Robotics and Automation
[32] Leonard J, Durrant-Whyte H. Simultaneous map building and localization for
Letters since 2015, and for IEEE Robotics and Automation Magazine since 2017.
an autonomous mobile robot. In: Intelligent robots and systems ’91. ’intel-
ligence for mechanical systems, Proceedings IROS ’91. IEEE/RSJ international
Cristian Secchi received the M.Sc. degree in Computer
workshop on; 1991. p. 1442–1447 vol.3.
Science Engineering, from the University of Bologna, Italy,
[33] Hornung A, Wurm K, Bennewitz M, Stachniss C, Burgard W. Octomap: an ef-
in 20 0 0 and the Ph.D. in Information Engineering, from
ficient probabilistic 3d mapping framework based on octrees. Auton Robots
the University of Modena and Reggio Emilia, Italy, in
2013;34(3):189–206.
2004. He has been visiting student at the University of
[34] Pietzsch S, Vu TD, Burlet J, Aycard O, Hackbarth T, Appenrodt N, et al. Results
Delft (NL) and at the University of Twente (NL) in 20 0 0
of a precrash application based on laser scanner and short-range radars. Trans
and in 2002 respectively. He is currently an Associate Pro-
Intell Transport Sys 2009;10(4):584–93. doi:10.1109/TITS.20 09.203230 0.
fessor at the University of Modena and Reggio Emilia.
[35] Thrun S. Learning occupancy grids with forward models. In: Intelligent robots
His research deals with human-robot physical interac-
and systems, 2001. Proceedings. 2001 IEEE/RSJ international conference on, 3;
tion, telerobotics, mobile robotics and surgical robotics
2001. p. 1676–1681 vol.3.
and he has published more than 100 papers on interna-
[36] Bar-Shalom Y, Campo L. The effect of the common process noise on
tional journals and conferences. His PhD thesis has been
the two-sensor fused-track covariance. Aerosp Electr Syst IEEE Trans
selected as one of he three finalists of the 5th Georges
1986;AES-22(6):803–5.
Giralt Award for the best PhD thesis on robotics in Europe. He has been co-chair of
[37] Bar-Shalom Y. On the track-to-track correlation problem. Autom Control IEEE
the IEEE Robotics and Automation Society Technical Committee on Telerobotics for
Trans 1981;26(2):571–2.
2007 to 2012. He has been a guest Editor for the Special Issue on Design and Con-
[38] Kalandros M, Trailovic L, Pao LY, Bar-Shalom Y. Tutorial on multisensor man-
trol Methodologies in Telerobotics, published on Elsevier Mechatronics in October
agement and fusion algorithms for target tracking. In: American control con-
2010. He has served as Associate Editor of the IEEE Robotics and Automation Mag-
ference, 2004. Proceedings of the 2004, 5; 2004. p. 4734–4748 vol.5.
azine from January 2005 to December 2008. Since September 2012 he is serving as
[39] Khawsuk W, Pao LY. Decorrelated state estimation for distributed tracking of
an Associate Editor of the IEEE Transactions on Robotics and since June 2015 he is
interacting targets in cluttered environments. In: American control conference,
serving as an Associate Editor for the IEEE Robotics and Automation Letters.
2002. Proceedings of the 2002, 2; 2002. p. 899–904 vol.2.
[40] Xu L, Krzyzak A, Suen C. Methods of combining multiple classifiers and
Cesare Fantuzzi Received the M.S. degree in Electrical En-
their applications to handwriting recognition. Syst Man Cybern IEEE Trans
gineering and Ph.D. degree in System Engineering from
1992;22(3):418–35.
the University of Bologna (Italy) in 1990 and 1995, re-
[41] Kuncheva LI, Bezdek JC, Duin RP. Decision templates for multiple classifier fu-
spectively. From 1996 to 20 0 0 Dr. Fantuzzi was an Assis-
sion: an experimental comparison. Pattern Recognit 2001;34(2):299–314.
tant Professor of Automatic Control at University of Fer-
rara (Italy), from 20 0 0 to 20 06 he was an Associate Pro-
Elena Cardarelli is a Postdoctoral Researcher at the De-
fessor of Automatic Control at University of Modena and
partment of Sciences and Methods for Engineering, Uni-
Reggio Emilia, Engineering Faculty in Reggio Emilia, and
versity of Modena and Reggio Emilia, Italy, since 2013.
from 2006 to now he is a Full Professor in the same Fac-
She received the Dr. Ing. (Master) degree in Computer
ulty. He has held a variety of positions as project leader
Engineering in 2007 and the Ph.D. degree in Informa-
in several applied research programs developed in coop-
tion Technology in 2012, both from the University of
eration with small, large and multinational companies in
Parma, Italy. From 2007 to 2013 she has been research
the field of packaging machinery builder and manufactur-
Fellow at the Artificial Vision and Intelligent System
ing sector. His research interests are principally in the theory and application of
Laboratory (Vislab) at the Dipartimento di Ingegneria
discrete-event control for automatic machinery and in the modeling and control of
dell’Informazione, University of Parma. Her main research
Mechatronics system and robots.
interests include computer vision, image processing, pat-
tern recognition and machine learning. He is participating
to the FP7 PAN-Robots (Plug and Navigate Robots) project.
From 2011 to 2013 she has partecipated to the FP7 V-Charge (Automated Valet Park-
ing and Charging for e-Mobility) project.

Valerio Digani is a R&D Scientist/Engineer at Elettric 80

spa. He received the Bachelor and the Master degree in
Mechatronic engineering from the University of Modena
and Reggio Emilia in 2010 and 2012 respectively, and his
Ph.D. in Innovation Industrial Engineering (Mechatronics
curriculum) from the University of Modena and Reggio
Emilia in 2016. Its Thesis was titled “Traffic Coordination
for AGV Systems: an Ensemble Modeling Approach”. In
2014, he has been a Visiting Researcher at Drexel Univer-
sity, Philadelphia, Pennsylvania (USA), for 6 months. The
main research interests include multi-robot system, coor-
dination and path planning and optimization algorithms.
He currently works on predictive traffic models, task as-
signment problems, path planning and data fusion algorithms for industrial envi-
ronments (AGV Systems). He has participated to the European FP7 PAN-ROBOTS
(Plug and Navigate Robots) project and to the 2011/2012 Sugar project in collab-
oration with Stanford University (USA), HSG University (Switzerland), University of
Modena and Reggio Emilia (Italy) and AUDI.