Content uploaded by Alessandro Palleschi
Author content
All content in this area was uploaded by Alessandro Palleschi on Oct 30, 2020
Content may be subject to copyright.
WRAPP-up: a Dual-Arm Robot for Intralogistics
Manolo Garabini∗,+, Danilo Caporale∗,+, Vinicio Tincani ],∗,
Alessandro Palleschi∗,+, Chiara Gabellieri ∗,+, Marco Gugliotta∗, Alessandro Settimi∗,+,
Manuel G. Catalano ],∗, Giorgio Grioli],∗, Lucia Pallottino∗,+
Abstract— The diffusion of the e-commerce has produced
larger and larger volumes of different items to be handled
in warehouses, with the effect to increase the need for pick-
ing automation. Conventionally, automation can be achieved
through a custom plant in case of large scale productions where
the items have well-known characteristics that are expected to
change slowly and little over time. However, today the challenge
is to realize a solution that is flexible enough to handle goods
with different shapes, sizes, physical properties, and grasping
modes. To solve this problem we first analyzed how humans
perform picking and then synthesized their behavior in four
main tactics. These have been used as guidelines for the design,
the planning and the control of WRAPP-up: a dual arm robot
composed of two anthropomorphic manipulators, a Pisa/IIT
SoftHand and a Velvet Tray. The system has been validated
and evaluated through extensive experimental tests.
I. INTRODUCTION
E-commerce, i.e., buying and selling physical goods via
services over the internet, has now reached his full devel-
opment. Led by Amazon, which accounted for more than
50% of the growth of the whole e-commerce market, and
by Alibaba, in 2017, retail e-commerce sales amounted to
more than 2 USD trillion with an annual growth rate higher
than 25% [1]. The expansion of e-commerce is affecting the
way warehouses work, especially the intralogistics, i.e., the
internal flow of goods within a distribution center [2].
On one side, the market growth led to an increase of the
employment: data from the Census Bureau [3] show that, in
the U.S., there was an annual growth rate in the Warehousing
and Storage (North American Industry Classification System
- NAIC - 493) employment of the 28% (from 2015 to 2016)
and that in 2016 the total workforce reached more than 600K
units. An analysis conducted by Data USA on the Census
Bureau ACS PUMS 1-Year Estimate data shows that material
movers are the largest share (20%) of jobs [4].
On the other side, a strong effort is devoted to maxi-
mize intralogistics efficiency by fully employing optimiza-
tion techniques [5], by pushing the productivity of human
operators even if it may cause high workloads [6], and by
adopting automated solutions. Finally, e-commerce impacted
both business-to-consumer (B2C) and business-to-business
(B2B) markets. The market size of B2B e-commerce is more
∗Centro di Ricerca “Enrico Piaggio”, Universit`
a di Pisa, Largo Lucio
Lazzarino 1, 56126 Pisa, Italy
+Dipartimento di Ingegneria dell’Informazione, Universit`
a di Pisa, Largo
Lucio Lazzarino 1, 56126 Pisa, Italy
]Soft Robotics for Human Cooperation and Rehabilitation, Fondazione
Istituto Italiano di Tecnologia, via Morego, 30, 16163 Genova, Italy
manolo.garabini@unipi.it
Fig. 1. WRAPP-up: a dual arm robot composed of two anthropo-
morphic 7-dof manipulators, a Pisa/IIT SoftHand, and a the Velvet
Tray. In the picture WRAPP-up is picking a box that does not have
the top surface.
than ten times the one of B2C [7], allowing to show an
unprecedented variety of products to the customers. This
brought undeniable advantages in terms of sales [7] but
also increased the flexibility requirements with which the
intralogistics system must comply.
Order picking - the process of retrieving products from
storage (or buffer areas) in response to a specific customer
request - is responsible for the 50-75% of the total costs
for a conventional warehouse [8]. Hence, order picking is
considered one of the highest priority areas to improve to
maximize a warehouse productivity. However, despite the
crucial importance of picking operations in warehouses, they
still mostly rely on human workers [9].
The major challenge preventing the full automation of
picking is represented by the high variability of objects to
handle in terms of shapes or object configurations, not ac-
cessible or even absent surfaces, flexible or pierced surfaces,
to name a few. Regarding object shapes, cuboids constitute
the vast majority of all the items stored in warehouses [10].
According to [11], among the shipped packages, the shapes
that occur more often are cuboids and, even if in a lower per-
centage, cylinders. Thus, strategies to manipulate cylinders
and cuboids in different configurations handle a considerably
large part of the goods in intralogistics processes.
Several are the components that concur to the realiza-
tion of a flexible picking solution: a robot design able to
execute the picking operations in a warehouse environment
physically; a vision system to detect objects and constraints;
a perception system able to identify desired and undesired
contacts; a planning method able to generate a trajectory
accomplishing the task while satisfying the constraints and
adapt it based on the perception outcomes; a control strategy
able to track the desired trajectory. This work is focused on
the realization of a flexible, autonomous picking solution,
leaving for future works the integration with the vision
system.
Some of the most challenging and common situations an
autonomous Pick and Place system might encounter are: 1)
reduced collision-free end-effector poses due to other goods
or containers; 2) restricted portion of the external surface of
the object available for gripper contact due to other goods
especially when tightly packed together; 3) deformability of
the object, meaning that the shape of the grasped objects
changes under external forces; 4) porosity of the object,
which prevents the employment of simple and nimble suction
grippers. Despite the great effort in the development of
picking solutions, as far as the authors know, no existing
automatic solution is flexible enough to cope with such
challenges, many of which may occur together. A detailed
problem statement is reported in Sec. III, while an overview
of the existing solutions on the market is reported in Sec. II.
The main contributions of this work are the design, realiza-
tion, and testing of WRAPP-up, a novel human-inspired dual
arm robot for intralogistics. The development of WRAPP-
up relies upon observing the techniques adopted by human
pickers at work in warehouses. Indeed, by observing expert
operators, we identified four main maneuvers they commonly
adopt, which are detailed in Sec. IV. Based on these findings,
we designed a dual arm robot (see Sec. Vfor more details)
composed of two 7 degrees-of-freedom manipulators and two
different end-effectors: an adaptive end-effector able both
to grasp a large variety of objects and to stably interact
with different shapes, and a tray with an actuated belt.
Moreover, we encoded the human observed picking strategies
into parametric motion primitives adopted in the trajectory
planning of the robot. Finally, an extensive experimental
validation (see Sec. VII) has been conducted.
To the best of the authors’ knowledge, WRAPP-up is
the first autonomous picking system able to approach the
whole spectrum of picking tasks of the intralogistics: from
bin picking to pallet picking.
II. REL ATE D WORK
Picking tasks can be classified based on different parame-
ters, of which one of the most important is the location from
which the items should be grasped.
On one side, there is the bin-picking problem in which the
objects are typically placed either in an ordered way (or not)
in a box and a single object has to be grasped. Often the
object has size and weight such that it can be handled with
one end-effector. The problem of grasping a single object
with an ad-hoc end-effector has been extensively studied
from theoretical and experimental viewpoints. However, bin-
picking still represents an open challenge, especially in un-
structured environments. This is also testified by challenges
aimed to enhance the warehouse automation in picking
operations, such as the Amazon Picking Challenge [12]. The
interested reader is referred to [13] for a comprehensive
review of robotic picking and to [14] and [15] for recent
results.
On the other side, there is the picking of items that are
located on a pallet. To automate this task two main different
solutions can be adopted: a mobile manipulation approach
in which the robot is provided with a mobile base, or a
grounded manipulation approach in which the pallet (or the
shelf) is brought to the manipulator by mobile devices [16].
Prominent examples of autonomous mobile manipulation
platforms for logistics include: Little Helper III [17], DLR
omniRob [18], and Handle [19]. The first two robots are
mainly devoted to picking objects from shelves. They consist
of a robot arm with a two-fingered parallel gripper mounted
on a stable mobile base. Handle instead has an unstable
two-wheeled mobile base, which requires a more expensive
control but substantially reduces the robot’s footprint, is
equipped with a vacuum gripper, and is devoted to box
handling. The interested reader is referred to [20] for com-
prehensive literature reviews on the subject. Furthermore,
Magazino [21] and InVia Robotics [22] sell two products
on the market, mainly devoted to box picking and based on
suction cups. Both the solutions exploit a picking strategy
based on box sliding, which may not be suitable for boxes
stacked one upon the other and, in general, not free to slide.
TORU, the robot by Magazino, is suitable for picking small
boxes from shelves, especially shoe boxes. It has also been
integrated with a different picking strategy, always based on
objects sliding, and additionally requiring the accessibility of
the rear surface of the box [23].
A recent example of a grounded manipulation approach
is the Dora Picker [24]. Its novelty relies on the soft and
adaptive design of the end-effector. The ground-based depal-
letizing robots available in the market, despite their different
working principles, share the drawback of being bulky and
not easily relocatable [25]–[27]. Moreover, unlike WRAPP-
up, they are usually suitable solely for pallet picking, while
a different robot would be necessary for bin-picking. See,
e.g., the example of Swisslog [28], which proposes on
its website two robotic solutions, one for picking larger
boxes, ACPaQ, and one for bin-picking, ItemPiQ. Another
key distinguishing aspect of WRAPP-up compared to the
integrated solutions proposed by Swisslog is that the former
is aimed to incorporate in a unique platform both picking
and discharging.
Currently, the most flexible autonomous systems that may
be used to accomplish picking tasks are provided with
mechanical and vacuum end-effectors. A wide overview of
the most recent development in gripping devices can be
found in [29].
Mechanical end-effectors for prehensile tasks are certainly
the most widespread and many of them fall into two neatly
distinct categories: simple grippers [30], [31] and complex
or anthropomorphic hands [32]–[34].
Among them, several devices exist that trade simplicity for
flexibility. Examples include underactuated and soft grippers
[35]–[37] and end-effectors with active surfaces. An example
of a versatile mechanical gripper is the Traction Gripper
[38]. It has a shaped frame with counter-rotating belts, which
exploits the friction forces to pull the boxes towards the
corners of the frame and firmly hold them in position. An
evolution of this concept is the Fraunhofer Roll-on Gripper
[39], which is a hybrid between a lift and a gripper. In
this solution, the belts also allow for manipulating (translate
or rotate) the boxes once picked up by the gripper. A
similar solution is exploited by Premium Robotics [25].
These grippers alone show limitations when the grasped item
is lodged in the object below.
Vacuum grippers are widely employed for the grasping
of boxes by their top surface ( [26], [27]), e.g., [40]. The
vacuum gripper by Wynright Robotics [41] is able, as well
as others ( [21], [22]) to grasp from the frontal side. The
device relies on an array of vacuum cups to drag the box on
a support surface of the gripper.
Vacuum grippers show severe drawbacks when the inter-
ested surface of the object is not suitable to be grasped, e.g.,
because the top surface is not present at all, or it is not robust
enough to sustain the weight of the object.
III. PROB LEM DE FINI TIO N
The task of interest, in this work, consists of picking
several different goods from single-item pallets, namely
pallets composed of several units of an item. The input to
the picking system is the sequence of goods to be picked
and their location on the pallet that may be provided by
a vision system. The design of a robot for picking tasks
depends on the size and the shape of the objects that must
be manipulated, and on the modes that can be profitably used
to grasp them. Picking tasks, the vast majority of which is
currently executed by human operators, can be classified into
two main categories: the ones that can be accomplished with
a hand and the ones that must be accomplished with two
hands. The problem of picking objects with a single end-
effector has been extensively treated in the literature and
in previous works of the authors [42] and [15]. This work
is focused on the problem of picking objects that humans
cannot pick with one hand. Notably, the solution that will
be proposed in this work will be able to accomplish both
categories of picking tasks. In the following, a list of items
that represents the 40% of the volume of a food warehouse
is reported together with their main features. Notice that the
food & beverage is one of the market segments most affected
by the e-commerce revolution, which today allows customers
to have their shopping bags directly delivered at home. In the
next subsections, the functional requirements that a picking
system should satisfy to work in such a warehouse profitably
are stated. Finally, the main challenges to be tackled in the
design and realization of this device are described.
A. Objects
The complete list of objects to be picked with their size
and weight is reported in Table II. The objects can be
grouped into two sets, depending on their shape: boxes or
cylinders. For the boxes, the values for the length (L), height
(H), and width (W) expressed in centimeters are reported,
while for the cylinders, the values for the diameter (D) and
the height (H) are listed.
B. Functional Requirements
A list of functional requirements that a picking system
should match is reported in Table I. It is important to stress
that they should not be taken as absolute values for every
intralogistics company but for operators that manage a set
of objects comparable to the one reported in Table II. These
are grouped into key performance areas and indicators. Their
quantitative value should be considered as a target for the
picking system.
TABLE I
PICKING TASK TARGET PERFORMANCE
Performance Area Performance Indicator Target Unit
Productivity Average time to empty a pallet 123 min
Picks per hour 180 #
Reliability First-Attempt Success Rate 90 %
Productivity performance indicators have been computed
based on the fact that such a robot would be economically
sustainable if it is able to perform 3 picking movements every
minute (corresponding to 180 picks per hour). Given that
a Euro-pallet (80 x 120 cm base piled up to 1.5 meters)
may contain up to 627 smallest items, and up to 115 largest
items (among the objects reported in Table II), the time
needed to empty a pallet can be evaluated in 209 and 38
minutes, respectively. The average of these two values gives
the productivity performance indicator reported in the table.
The picking success rate takes into account the grasping
system without considering the vision system.
C. Challenges
The main challenges of the picking phase can be identified
in the following 3 points:
•boxes are often very close to each other, with two op-
posite sides, which are the most desirable for a reliable
and robust grasp, usually not easily accessible. Hence,
to be properly handled, they should be first moved to
TABLE II
OBJECTS CONSIDERED FOR PICKING TASKS,WITH THEIR WEIGHT (KG), S IZE (C M), AN D REL ATE D GRASPIN G STRATE GIES
9.1kg - 22×26 4.3kg - 20×25 3.2kg - 16×18 3kg - 20×13
Horizontal Rotation Horizontal Rotation Horizontal Rotation Horizontal Rotation
3kg - 19×29×14 2.7kg - 40×7×30 2kg - 40×8×30 1.8kg - 15×13
Horizontal Rotation Horizontal Rotation Horizontal Rotation Horizontal Rotation
9kg - 46×15×18 6.0kg - 40×10×26 5kg - 46×18×15 2.5kg - 38×18×14
Vertical Rotation - Sliding Vertical Rotation - Sliding Vertical Rotation - Sliding Vertical Rotation - Sliding
2.5kg - 36×15×23 2.0kg - 38×23×20 1.9kg - 35×13×16 0.4kg - 40×18×25
Vertical Rotation - Sliding Vertical Rotation - Sliding Vertical Rotation - Sliding Vertical Rotation - Sliding
(a) Human grasping a box–shaped object.
(b) Human grasping a cylinder–shaped object.
(c) Human grasping a cylinder–shaped object.
(d) Human grasping a box–shaped object.
(e) Human grasping a box–shaped object.
Fig. 2. Human grasping different shaped objects with different strategies.
guarantee that two opposite faces are accessible and then
picked.
•some of the items do not have a top surface, or it may
not be suitable to grasp the object. These objects cannot
be grasped with vacuum grippers.
•the bottom side of some objects is recessed in the upper
side of the objects which lie under, or more in general,
they can not slide. This means that the objects can only
translate along the vertical direction or rotate about a
horizontal axis.
IV. HUMAN PICKING SKILLS
There is not a systematic method to synthesize a system
that matches all the requirements listed in Sec. III, and one
of the reasons is that it simultaneously involves the co-
design of the robot structure, planning, and control. Hence
we observed skilled human operators at a food warehouse
during the execution of manipulation tasks when picking the
objects listed in Tab. II. More in detail, we recorded two
human operators from a food warehouse while performing
the picking action. Each picking action has been repeated
three times. These live observations and the analysis of the
video recordings led to two lessons learned:
•bi-manual manipulation has a crucial role in picking
operations since humans use both hands to manipulate
and handle the objects. In the majority of the tasks, one
hand is used to move the object, and the other hand is
used as a support;
•the strategies to pick the items are classifiable in three
main classes.
The strategies, shown in Fig. 2, that human operators adopt
while manipulating goods fall into 3 groups depending on
object shape and form factor:
a) Rotation about Horizontal Axis: In case of thin
boxes, i.e., H>W,H>L, cylindrical objects, or if the
support surface of an object cannot slide, the operators use
one hand to rotate the goods about a horizontal axis and to
put the object on the supporting hand (see Fig. 2(a,b,c) and
Fig. 2(d)).
b) Rotation about Vertical Axis: For thick boxes (H<
W,H<L) with no constraints at the base, the horizontal
rotation is not convenient because of the less favorable lever
arm; thus the operators decide to rotate the boxes about a
vertical axis to have access to the back surface of the object,
as in Fig. 2(d). This strategy can then evolve in two different
picking continuations. In the first one, the box is picked up
by two opposite surfaces while the operator uses his hands
like the jaws of a parallel gripping device. In the second
one, the box is first dragged towards the worker acting on
the back surface and then supported by the other hand as the
box sticks out the pallet or the underneath layer of goods.
This helps apply two different grasping strategies: grasp the
object relying on contacts on two opposite surfaces (front
and back) and slide the object relying on the contact on the
back surface.
c) Sliding: For thick boxes with no constraints at the
base, the operators push or pull the objects until they reach
the support hand at the boundary of the pallet, as in Fig. 2(f).
Picking strategies that human operators adopt highlight
that often, pickers naturally choose different functions for
each hand. One hand is mainly used to accomplish manip-
ulation tasks: pushing an object in the sliding strategy and
adapting to the shapes of the different objects in the other two
strategies. The other hand is often responsible for supporting
the majority of the item weight and may be used to perform
placing operations such as alignment and unloading.
V. WRAPP-UP DESIGN
Inspired by the techniques adopted by warehouse workers,
the envisioned solution is a dual arm system. The system is
composed of two Lightweight Robots arranged to perform
pick and place tasks properly. The mounting bases of the
arms are fixed at an established relative pose, as better
described in the following. Two different end-effectors, pro-
vided with six-axis force/torque sensors, are attached to the
wrists of the robot arms. To perform the dexterous operation,
one arm is featured with a Pisa/IIT SoftHand: a human-
like, adaptive, robust artificial hand the closure movement
of which is easy to control since it is actuated by a single
motor. The second end-effector is the Velvet Tray, which
serves as a support tool. A brief description of the design of
the two end-effectors is provided at the end of this section.
In our preliminary experimental set up each robotic arm
has been mounted on an independent movable gate, which
allows for three DOFs in a plane (two translations and
a rotation), enabling configurable relative locations of the
two arms (see Fig. 3(c)). With this test-bench, we could
easily test different relative positions of the arms in a store-
like environment. The choice of the most suitable relative
configuration of the arms is described in the next subsection.
In a final use case, the dual arm robot can be mounted on a
fixed base as well as on a mobile base, depending on the end-
user requirements. For instance, the end-user may employ a
picker-to-goods strategy, thus needing a mobile base for the
picking station, or a goods-to-picker strategy, in which the
picking station is fixed and the objects are brought there. At
the moment, our robot is mounted on a fixed base, but it
will be integrated on an autonomous mobile robot in future
works.
A. Relative pose of the two manipulators
Once the overall structure of the robot, the end-effectors,
and the manipulation strategies have been defined, the robot
design can be detailed. Particularly important is the location
of the two arms w.r.t. each other and the pallet. A wrong
relative location of the arms may prevent the correct execu-
tion of a strategy because of two main reasons: one joint (or
more) reaches the limit of its range of motion, or a point of
the desired trajectory is out of the reachable workspace of
the bimanual system.
To properly choose the relative location of the arms, a two-
step strategy has been adopted. First, a manipulability index
for each arm has been evaluated for a set of points to find
a relative pose that provides an adequate superposition of
the manipulators’ dexterous workspaces. Then, a feasibility
analysis, conducted by simulating the kinematic execution
of the robot trajectories (based on the strategies presented in
Sec. VI for the objects listed in Tab. II), has been performed
to check that the relative pose found in the first step allowed
the robot to operate at least over half of the pallet footprint.
The feasibility phase took into account constraints due to
joint limits and realistic external obstacles, e.g., floor and
shelves. It is worth pointing out that, for the first step of the
strategy, other metrics could have also been implemented,
e.g., to include the direction of maximum force of the
arms. Still, at this stage, we preferred to give priority to
the manipulability. Future works will be devoted to the
evaluation of different metrics. Among the manipulability
(a) An example of the manipulability analysis showing the manipulability
measure and the preferred directions of motion.
(b) Configuration of the two arms in
the current setup. ∆x=20cm, ∆z=
30cm, ∆θ=15deg (θR=45 deg and
θL=60deg).
(c) CAD Model of the experimental
setup
Fig. 3. Analysis of the relative configuration of the manipulators, resulting relative configuration, and CAD Model of the experimental setup.
TABLE III
MAN IPUL ABI LITY A NALYS IS FOR D IFFE RENT CO NFI GU RATIO NS O F T HE T WO A RM S
Workspace Volume - VU(m3) 5.91 7.13 6.43
Shared Workspace Volume - VI(m3) 2.60 1.38 1.23
Intersection Average Manipulability ¯wD|I0.43 0.46 0.47
Manipulability M0.19 0.09 0.09
measures suitable to quantify the ability of a robot to execute
a movement in an arbitrary direction of the Cartesian space,
from a given pose q, we use the one presented in [43]
w(q) = qdet(J(q)J(q)T),(1)
that is a measure of the volume of a six-dimensional ellipsoid
of which the semi-axes length is represented by the square
roots of the singular values of the end-effector Jacobian J(q).
The eigenvector of J(q)corresponding to the largest singular
value represents the easiest direction of motion. Fig. 3(a)
represents an example of a graphic result that shows the
manipulability index w(q)(evaluated according to Eq. 1) and
the preferred directions of motion of the two robots.
Given a relative location of the arms, we evaluate the
manipulability of the configuration using a scaled average
manipulability for the two arms [44] and the volume of the
shared workspace. The average manipulability of each arm is
evaluated by averaging the manipulability index values at N
uniformly-sampled feasible configurations in the joint space.
The maximum value manipulability index then scales this
value according to
¯wi=∑N
j=1wi(qj)
maxj{wi(qj)}N.(2)
Then, we defined the manipulability index for the dual arm
(a) The Pisa/IIT SoftHand is the dex-
terous end effector of WRAPP-up.
Furthermore, the picture shows also
all the components attached to the
wrist of the robotic arm.
(b) The Velvet Tray is the support end effector of Wrapp-up. On the left a 3D view is shown. On the right a cut
view shows the power transmission group and the conveyor belt.
Fig. 4. The Manipulation End-Effector (Pisa/IIT SoftHand) and the Support End-Effector (Velvet Tray).
system as
¯wD=¯w1+¯w2
2(3)
where ¯w1and ¯w2are the average manipulability indexes for
the first and second arm, respectively. Thus, the configuration
manipulability is then expressed as
M=VI
VU
¯wD|I(4)
where we defined with VUthe workspace of the dual arm
system (obtained by the union of the workspaces of the two
manipulators), with VIthe volume of the shared workspace,
and with ¯wD|Iwe denote the average manipulability index
for the dual arm system (defined as in (3)) computed using
only the configurations belonging to the intersection of
the workspaces of the two manipulators. The poses with
high manipulability are the ones that give large dexterous
collaborative workspaces.
The solution provided by the manipulability analysis is
a reasonable starting point. However, its selection does not
take into account the tasks that the robot should execute.
To evaluate the quality of the selected configuration for our
task, we simulated the execution of picking tasks using the
strategies described in Sec. IV for different positions of the
target object and registered the associated Cartesian error. To
perform this analysis, we considered the object as they were
placed on a 0.8m×0.6m×1.5m pallet in front of the robotic
platform so that it would lie within its reachable workspace.
The width corresponds to half of the width of a EURO-pallet.
We simulated the task for every poses an object could assume
on this reference pallet (the possible poses are limited and
depend on the shape and the dimension of the object), and
we checked that, in the selected configuration, the robot was
able to execute the task with a limited Cartesian error (under
1cm of error).
To reduce the complexity of the search of the relative
configuration, we predefined a set of reasonable candidate
poses and evaluated the manipulability for each of them and
evaluated the task performance for the most promising one.
Examples of the manipulability for three different candidate
configurations are reported in Tab. III, where the values have
been obtained using N=20000 samples. It is worth noting as
the first configuration (the one we have eventually selected)
is the one with the higher manipulability. A detail of the
manipulators’ base relative location is depicted in Fig. 3(b).
B. Mechanical Design of the Manipulation End-effector
The manipulation end-effector is the Pisa/IIT SoftHand
(Fig. 4(a)). Its mechanical robustness and adaptability, to-
gether with the ease of control, make it particularly suitable
for the type of use required to accomplish the picking task.
For an in-depth description of the hand the reader can refer to
[45]. The Pisa/IIT SoftHand (7) is attached to the wrist flange
of the robotic arm (1) through a 6-axes Force/Torque ATI-
Mini45 sensor (2) and four rubber beams (6). The ATI sensor
detects changes in the state of the hand, such as contact
with the objects to be manipulated in regular functioning but
also undesired collisions, preventing the end-effector from
damaging. The rubber beams are located between the end-
effector and the ATI-sensor. They favor the slowdown of the
external force loading rate in case of collision, increasing the
time for a rapid emergency-stop response. Toothed flange (5)
crimps plate (3) (fixed to the sensor) and plate (4) (fixed to
the hand side) together.
C. Mechanical Design of the Support End-effector
The end-effector that functions as a support for the goods
to be manipulated is the Velvet Tray. Fig. 4(b) shows a
3-D model of the Velvet Tray. Its design and principle of
operation are inspired by research carried out by the Authors
about grippers with active surfaces [46]–[48]. It is equipped
with an actuated belt to ease the loading maneuvers of
the goods. It is attached to flange wrist (1) of the KUKA
arm through flange (2). Between the KUKA arm and the
Velvet Tray, a 6-axes Force/Torque ATI-Mini58 sensor (3)
and rubber beams (4) are interposed with the same aim as in
(a) Strategy to grasp an object from behind (side view). (b) Strategy to roll a tall object (side view).
(c) Strategy to rotate an object exploiting environmental constraints, such as
another object used as a pivot point (top view).
(d) Strategy to lift an object and put it on the tray (side view).
Fig. 5. Robot grasping strategies.
the Pisa/IIT SoftHand. The elastic junction is here composed
of 10 rubber beams (4) arranged in a circle, which slow down
the loading rate of the external forces in collision events. Belt
(8) is coated with high grip polyurethane with a coefficient
of static friction polyurethane-steel of 0.8 which allows an
inclination of 38◦with respect to a horizontal plane without
a mass on the belt sliding down. This value of the friction
coefficient provides the worst-case for the torque of the motor
that guarantees to hold the target mass of 2.5kg with the
Velvet inclined 38◦. A Maxon motor DCX22 actuates the
belt with gear-head GPX83 (5) that is able to move a mass
of 2.5kg with an inclination of 38◦within the continuous
functioning condition of the driver. The power transmission
between the motor and the driver roll of the belt is due to
gears (6). Tension roller (7) ensures a proper tension in the
belt of at least 20Nnecessary to transmit the required torque.
Finally, a set of idle rollers (9) sustain the objects and form
an approximately flat surface under the belt.
VI. TR A NS LATING HUMAN PICKING SK IL LS IN TO
ROB OT MOT ION PRIMITIVES
Inspired by the observation of the strategies adopted by the
operators (described in Sec. IV) parametric motion primitives
have been defined to plan the motion of the robot during the
task execution:
•Sliding one end-effector is used to push (or pull) an
object towards the other end-effector, which secures the
grasping and may support the weight of the object, as
shown in Fig. 5(a).
•Horizontal Rotation one end-effector is used to tilt
the object about a horizontal axis gently. This can be
achieved in two different cases: i) about a horizontal
axis on the front-bottom edge of the bounding box
enveloping the object (see Fig. 5(b)); ii) about an axis on
the back-bottom edge of a bounding box enveloping the
object (see Fig. 5(d)). In the first case, the object rotation
will end when it lays on the support end-effector. This
strategy is intended to be used with objects (boxes,
cylinders) of which the height is the largest dimension.
In the second case, the object’s rotation will allow the
second end-effector to be placed under the object as
support.
•Vertical Rotation one end-effector approaches the ob-
ject’s side, then rotates it about a vertical axis ideally
located at one edge of the object. Once the rotation
has produced enough room for the end-effector, it slides
inside this gap and proceeds to slide the object towards
the pallet’s exterior. This strategy is suitable when the
objects are compactly packed (see Fig. 5(c)), and it is
necessary to make room for the end-effectors to perform
a successful grasp.
For each object, the choice of the strategy has been made
based on its shape and on how the objects are stacked on the
pallet. Each motion primitive is defined as a set of Cartesian
waypoints for the two end-effectors, expressed w.r.t. a frame
placed on the object. The definition of waypoints that allows
for a correct manipulation of the object, e.g., to rotate or
tilt it in order to produce enough room for positioning an
end-effector as in 5(c) or 5(d), is the result of simulations
and real experiments on the objects. Thus, they depend on the
physical properties of the end-effectors and the objects. Once
the pose of the object is retrieved, and the correct primitive
is selected, the waypoints expressed in the object-fixed frame
are transformed in the world frame.
In order to take into account the robot kinematics and the
joint limits for the motion planning, we determined the path
at the joint level via the reverse priority algorithm described
in [49], which allows us to define a set of tasks with different
priorities including unilateral constraints (e.g., joint-position
limits). More in detail, for each of the two arms, we have set
the Cartesian pose tracking, i.e., the position and orientation
of the end-effector, as the low priority task, and the joint-
(a) Reactive planning architecture. The picking strategy is divided into
basic consecutive phases planned online. The information from the sen-
sors are used to detect contacts between the robot and the environment,
or if the end-effectors reached their target position. Depending on the
typology of the picking phase (Approach or Manipulation), these events
trigger the transition to the next phase or an emergency state.
(b) Force sensors readings for the two end-effectors during the strategy
execution. The transitions between the states into which the strategy is
decomposed are highlighted.
(c) Sequence of a picking strategy in which each block corresponds to a state of the state machine on which the planner is built.
Fig. 6. Reactive Planning Approach.
position constraints as the high priority tasks. Then the path
following could be achieved by the minimum-time approach
presented in [50].
Accounting for realistically imperfect knowledge of the
object position provided by a vision system in a future
integration, the force/torque sensors have been exploited
to plan the trajectories based on contacts with the objects
reactively. Indeed, the measured forces can be used to detect
possible contact with an object whenever they exceed a user-
defined threshold. Our reactive planning approach (see Fig.
6(a) for a schematic representation of the architecture) is
accomplished by decomposing each picking strategy into
consecutive basic phases represented by states of a finite
state machine. Each phase is planned online (block ”Plan
Picking Phase” of Fig. 6(a)) and generates the Cartesian
trajectory for the end-effectors. For each of the phases, the
kinematic feasibility of the planned trajectory is checked
(block ”Feasible” in Fig. 6(a)). In this block, the generation
of a path for the joints of the robot is performed using
the reverse priority algorithm. If the desired motion is not
feasible because, e.g., of the constraints on the joint ranges,
the task is aborted, allowing the intervention of a human
operator. The transition between a phase and the following
is triggered online based on the information coming from
the force/torque sensors and the joint position sensors of the
robot. This information can be used to detect two possible
events (block ”Event” in Fig. 6(a)): a detected contact (or
the loss of contact) or the end-effectors reaching their target
position. The reaction of the system at these events depends
on the typology of the picking phase. Indeed, the phases are
of two types: Approaching phases and Manipulation phases.
Approaching phases are the ones in which the end-effectors
have to establish contact with the object. This is a critical step
of the manipulation process, since an incorrect positioning
of the end-effector w.r.t. the object, possibly due to errors
on the estimate of the object pose, could cause the picking
failure. Hence the end-effector starts moving towards the
object along a specified direction until contact is detected.
Then, it stops, and the end-effector position at the contact is
used to update the object pose estimate. This refined estimate
is used to update the planned trajectory for the following
phases. If the expected contact does not happen within a
certain region, the system enters an emergency state, and
eventually, a human operator will be alerted. Conversely,
Manipulation phases are the ones for which a contact is
already established, and the end-effectors are manipulating
the object. The condition used to trigger the transition to
the successive phase is defined based on the end-effector
reaching the target position. Sensing of unexpected forces,
causes the system to enter an emergency state and eventually
alert a human operator, regardless of the specific phase. At
the end of the picking strategy, since the object will be placed
on the velvet tray, the force measurements can be used to
identify whether the object has been picked or if it fell.
(a) Horizontal Rotation Primitive.
(b) Cylindrical objects being picked using the horizontal rotation strategy.
Fig. 7. WRAPP-up picking cylindrical objects.
An example of the described reactive approach is reported
in Fig. 6(c) and 6(b), where the states of an example trajec-
tory and the corresponding force sensor readings are shown.
For this example, we considered a worst-case scenario where
the uncertainties on the pose estimate are such that require
additional pose refinement steps. Indeed, in phase (A) the
hand is approaching the object laterally to refine its pose. As
reported in Fig. 6(b), when contact is correctly detected (the
magnitude of the force Fxalong the contact direction exceeds
the threshold set at 10N), the pose along this direction is
updated, and the hand is placed front to the object and
starts moving towards it. Once again, the force measurements
inform about the established contact, and the robots enter
the third phase, (C), of the manipulation. In this case, since
a Horizontal Rotation is used, the hand lifts the bucket to
create space for the velvet tray. Therefore, the transition
toward the next phase is triggered by the hand reaching the
target Cartesian pose, and no force information is required.
Note that, to increase the robustness of the system, the force
measurements in this phase could detect the loss of contact
between the hand and the object and be used to abort the
current picking action. Regarding the other phases, contact
information is again used to trigger the transition from (D)
to (E) (the velvet tray is placed under the object, in contact
with it), and (E) to (F), where the hand is approaching the
object from the top to perform a collaborative sliding.
As said, the example shows the effectiveness of the
reactive approach even in case of a not precise knowledge
of the pose, which although requires the execution of the
redundant and time-consuming lateral approach (phase (A)).
Depending on the level of uncertainties on the pose estimated
by the perception system, such redundant steps could be not
necessary.
VII. EXP ER IME NTAL VALIDATION
In this section, the preliminary experimental validation
of WRAPP-up is reported. The picking capabilities of the
system are shown on a representative set of objects from the
ones in Tab. II. Indeed, the objects used for the tests allow us
to cover the two main shapes we have identified to be relevant
for logistics, i.e., cuboids and cylinders. Furthermore, they
allow us to test all the four motion primitives we presented
(a) Horizontal Rotation Primitive.
(b) Two rows (8 pieces) of thin boxes being picked using the horizontal rotation strategy.
Fig. 8. WRAPP-up picking thin boxes.
and validate the platform’s performance in different picking
scenarios.
Fig. 7(a) shows the strategy used to manipulate objects
which have a characteristic cylindrical shape, better suited
for a horizontal rotation strategy. With this approach, the
hand is placed in front of the bucket and grasps its edge
allowing it to lift and tilt it. This movement allows the tray
to be placed beneath it as a support. Once the tray has been
correctly positioned, the hand can release the object, see e.g.
the last frame of Fig. 7(a), and the tray is used to collect and
deploy the bucket. In this case, the hand can be employed
to ease the tray during the picking phase.
An example of the described approach used to collect three
rows of objects placed on a pallet is shown in Fig. 7(b).
The horizontal rotation strategy is also effective to pick
thin boxes, see Fig. 8(b).
In this case, the hand is placed behind the box and tilts it
until the box lays on the tray placed in front of the object
with a proper inclination. Then, the hand is used to ease
the object picking, keeping it on the tray while the latter
is returning parallel to the horizontal plane. The former
approach has been tested to successfully pick 8 boxes close
to each other, as in Fig. 8(b), showing the robustness of
the designed strategy even in the presence of other objects
behind the handled box.
The best picking strategy is not chosen solely based on
the object’s shape, but it also depends on the location of the
object on the pallet and the position of the other possible
items. To show this concept, two different picking tests have
been performed on the same object (with a box-like shape)
depending on its different orientation, see Fig. 9and Fig. VII.
In the first test, the boxes are easily picked using a sliding
approach due to their configuration. The hand is placed
behind the box and used to pull the object towards the
tray. The situation is different and requires a more complex
strategy; if the boxes are in a different configuration, e.g.,
they are rotated by 90 degrees around the vertical axis w.r.t.
the previous case, and they are compactly packed, as shown
in Fig. VII.
This condition requires the use of a vertical rotation
strategy, where the hand approaches the box’s side and, eased
by the tray, rotates it about a vertical axis located at one edge.
Fig. 9. WRAPP-up picking thick boxes using sliding primitive.
Then, the hand slides inside the created gap and proceeds to
slide the box towards the tray.
Table IV shows the time for picking every single object
during the performed experiments. Then, an estimation of the
time to empty an entire pallet full of that object is reported.
To estimate the total number of boxes that are contained
in the pallet, the standard EU Pallet dimensions have been
considered for the base of the pallet, and a full pallet has
been considered to be 1.5 meters height. To compute the
number of objects that can be contained in such a pallet,
the dimensions of the objects have been taken into account.
A pallet of thin boxes contains thus 192 items, a pallet of
cylinders 360 items, one of the thick boxes 176 items in the
first case, and 165 items in the other. Hence, the time to
empty a pallet has been estimated, multiplying the average
time to pick an object for the number of objects in the full
pallet. Table Vreports the global performance indicators we
obtained for the picking task. The time to empty a pallet has
been computed as the average of the values reported in Table
IV. Fifty picking actions have been performed for each case
in order to test the system and estimate the values of the
performance indicators. The success rate is the average of
the four cases.
A. Discussion
This set of experiments was aimed at verifying the ef-
fectiveness of the hardware and the picking strategies in an
unloading simulation of goods on a pallet. These experiments
demonstrated that the WRAPP-up robotic platform is suitable
to fulfill the picking tasks of goods stacked on a pallet. A
comparison with the performance requirements specified in
Tab. I, highlights that the reliability requirement is met, but
the productivity should be improved through optimization
techniques (not yet integrated) that will be the subject
of future work. The performance reported in Tab. Vare
expressed for the platform in its current setup, i.e., without
a perception system and the optimization module for the
two arms and for the set of objects we used for the tests.
Indeed, the reliability has to be intended as an upper bound
for the system, since it does not include the presence of a
perception system. For the average time to empty a pallet, it
will have to be evaluated for the fully integrated platform, to
quantify the impact of the time required by the perception
system to retrieve the pose of the objects, and the impact
the optimization module presented in [50] in the overall
performance. It is worth noting as, with the current setup,
some of the more burdensome objects in Tab. II could result
in being difficult to handle because they exceed the nominal
payload of the arms. Strategies and solutions to effectively
handle those objects will be investigated. A more specific
study on the robustness of the platform for different objects
and setups, to evaluate how different friction forces could
impact the reliability of our manipulations that involve the
exploitation of environmental constraints, will be subjects for
future works.
VIII. CONCLUSIONS AND FUTURE WOR KS
In this work, we addressed the problem of realizing a
proof-of-concept robot that is flexible enough to manipulate a
variety of goods relevant to the intralogistics of warehouses.
Inspired by the picking strategies that skilled human oper-
ators adopt in the execution of these tasks, we realized a
Fig. 10. WRAPP-up picking thick boxes using the vertical rotation primitive.
TABLE IV
PICKING PERFORMANCE INDICATORS FOR THE FOUR SCENARIOS.
Object
Picking time per
object
55s 83s 16s 82.5s
Time to empty a
pallet
176min 498min 47min 227min
TABLE V
PICKING PERFORMANCE INDICATORS
Performance Area Performance Indicator Current Unit
Productivity Average time to empty a pallet 237 min
Reliability Picking success 92.5 %
dual arm robot provided with a Pisa/IIT SoftHand and a
Velvet Tray. The first end-effector is adaptable; hence it is
used to establish stable grasps to rotate and slide goods with
various shapes, while the second end-effector is mainly used
to support the weight of the objects. The robot has been
experimentally validated in multiple picking actions on a set
of four different representative objects. Future works will
include, on one side, to provide the robot planning with a
high-level decision tool that is able to automatically generate
the right strategy to adopt on the basis of features of the
objects that can be detected by a vision system. On the other,
to adopt suitable feedback strategies based on vision, force
feedback, and tactile feedback in order to improve robot
reliability. Furthermore, the average picking time will be
minimized by adopting suitable optimization algorithms, and
the robot will be provided with a mobile base.
ACKNOWLEDGMENT
This work was supported in part by the European Union’s
Horizon 2020 research and innovation program as part of the
projects ILIAD (Grant no. 732737), and in part by the Italian
Ministry of Education and Research in the framework of the
CrossLab project (Departments of Excellence).
REFERENCES
[1] C. McNair, “Worldwide retail and ecommerce sales: emarketer’s
updated forecast and new mcommerce estimates for 2016–2021,”
Industry Report, eMarketing, 2018.
[2] N. Boysen et al., “Warehousing in the e-commerce era: A survey,”
European Journal of Operational Research, vol. 277, no. 2, pp. 396–
411, 2019.
[3] United States Census Bureau, “2015-2016 SUSB Employment
Change Data Tables,” [Online] Available: https://www2.census.
gov/programs-surveys/susb/tables/2016/us 4digitnaics emplchange
2015-2016.xlsx?#, [Accessed: Jul. 1, 2020], 2019.
[4] Data USA, “Warehousing and storage [website],” Avaliable at: https://
datausa.io/profile/naics/493/#tmap top occs, [Accessed: Jul. 1, 2020],
2017.
[5] K. J. Roodbergen et al., “Designing the layout structure of manual
order picking areas in warehouses,” IIE Transactions, vol. 40, no. 11,
pp. 1032–1045, 2008.
[6] J. Bloodworth, Hired Six Months Undercover in Low-Wage Britain.
Atlantic Books London, 2018.
[7] K. C. Laudon and C. Guercio Traver, E-commerce: business, technol-
ogy, society. Pearson, 2017.
[8] C. G. Petersen and G. Aase, “A comparison of picking, storage, and
routing policies in manual order picking,” International Journal of
Production Economics, vol. 92, no. 1, pp. 11–19, 2004.
[9] R. Krug et al., “The next step in robot commissioning: Autonomous
picking and palletizing,” IEEE Robotics and Automation Letters,
vol. 1, no. 1, pp. 546–553, 2016.
[10] J. Rudd, A Practical Guide to Logistics: An Introduction to Transport,
Warehousing, Trade and Distribution. Kogan Page Publishers, 2019.
[11] W. Echelmeyer et al., “Performance indicators for robotics systems in
logistics applications,” in IROS Workshop on Metrics and Methodolo-
gies for Autonomous Robot Teams in Logistics, 2011, p. 55.
[12] N. Correll et al., “Analysis and observations from the first amazon
picking challenge,” IEEE Transactions on Automation Science and
Engineering, vol. 15, no. 1, pp. 172–188, 2016.
[13] J. Bohg et al., “Data-driven grasp synthesis - a survey,” IEEE Trans-
actions on Robotics, vol. 30, no. 2, pp. 289–309, 2013.
[14] J. Mahler et al., “Dex-net 2.0: Deep learning to plan robust grasps with
synthetic point clouds and analytic grasp metrics,” Robotics: Science
and Systems (RSS), 2017.
[15] C. Gabellieri et al., “Grasp It Like a Pro: Grasp of Unknown Objects
With Robotic Hands Based on Skilled Human Expertise,” IEEE
Robotics and Automation Letters, vol. 5, no. 2, pp. 2808–2815, 2020.
[16] Grey Orange, “Ranger robot series,” [Online] Available: https://www.
greyorange.com, [Accessed: Jul. 1, 2020].
[17] O. Madsen et al., “Integration of mobile manipulators in an industrial
production,” Industrial Robot: An International Journal, vol. 42, no. 1,
pp. 11–18, 2015.
[18] A. D¨
omel et al., “Toward fully autonomous mobile manipulation for
industrial environments,” International Journal of Advanced Robotic
Systems, vol. 14, no. 4, p. 1729881417718588, 2017.
[19] Boston Dynamics, “Handle,” [Online] Available: https://www.
bostondynamics.com/handle, [Accessed: Jul. 1, 2020].
[20] M. Hvilshøj et al., “Autonomous industrial mobile manipulation
(AIMM): past, present and future,” Industrial Robot: An International
Journal, vol. 39, no. 2, pp. 120–135, 2012.
[21] Magazino, “Pick-by-Robot: warehouse automation with Magazino
at FIEGE Logistik,” [Online] Available: https://www.youtube.com/
watch?v=kj8NaHAoLJwfiege%20lo, [Accessed: Jul. 1, 2020], 2017.
[22] InVia, “Introducing inVia Robotics,” [Online] Available: https://www.
youtube.com/watch?v=sCE0a8625tE, [Accessed: Jul. 1, 2020], 2018.
[23] Magazino, “Advanced Robotics - mobile picking robots for ware-
houses by Magazino,” [Online] Available: https://www.youtube.com/
watch?v=1j-RhPENu88, [Accessed: Jul. 1, 2020], 2017.
[24] H. Zhang et al., “Dorapicker: An autonomous picking system for gen-
eral objects,” in 2016 IEEE International Conference on Automation
Science and Engineering (CASE). IEEE, 2016, pp. 721–726.
[25] Premium Robotics, “Robots for logistics tasks,” [Online] Available:
http://premium-robotics.com/index.php?id=6&L=1, [Accessed: Jul. 1,
2020].
[26] ABB, “ABB depalletizing solution,” [Online] Available: https:
//www.youtube.com/watch?time$ $continue=1&v=8MbC9IaNvJM&
feature=emb logo, [Accessed: Jul. 1, 2020], 2019.
[27] FANUC, “Robotic system uses three FANUC palletizing robots to
service nine production lines - sen-pack,” [Online] Available: https:
//www.youtube.com/watch?v= lH2snLVO7w, [Accessed: Jul. 1,2020],
2017.
[28] Swisslog, “Robot-based picking & palletizing for higher warehouse
performance,” [Online] Available: https://www.swisslog.com/en-us/
products-systems- solutions/picking-palletizing- order-fulfillment/
robot-based- robotics-fully- automated, [Accessed: Jul. 1, 2020].
[29] G. J. Monkman, S. Hesse, R. Steinmann, and H. Schunk, Robot
grippers. John Wiley & Sons, 2007.
[30] Schunk, “Schunk grippers,” [Online] Available: https://schunk.com/
de en/gripping-systems/category/gripping-systems/schunk-grippers/,
[Accessed: Jul. 1, 2020].
[31] Willow Garage, “2G ”Velo” Gripper,” [Online] Available: http://www.
willowgarage.com/velo2g, [Accessed: Jul. 1, 2020].
[32] R. O. Ambrose et al., “Robonaut: Nasa’s space humanoid,” IEEE
Intelligent Systems, vol. 15, no. 4, pp. 57–63, 2000.
[33] L. Birglen et al.,Underactuated robotic hands. Springer Verlag,
2008.
[34] M. Grebenstein et al., “The DLR hand arm system,” in IEEE Interna-
tional Conference on Robotics and Automation (ICRA). IEEE, 2011,
pp. 3175–3182.
[35] A. Bicchi and A. Marigo, “Dexterous grippers: Putting nonholonomy
to work for fine manipulation,” The International Journal of Robotics
Research, vol. 21, no. 5-6, pp. 427–442, 2002.
[36] S. Hirose and Y. Umetani, “The development of soft gripper for the
versatile robot hand,” Mechanism and machine theory, vol. 13, no. 3,
pp. 351–359, 1978.
[37] A. M. Dollar and R. D. Howe, “Designing robust robotic graspers for
unstructured environments,” in Workshop on Manipulation for Human
Environments, 2006.
[38] Fraunhofer, “Traction gripper,” [Online] Available: https:
//www.iml.fraunhofer.de/en/fields of activity/material-flow-systems/
machines and facilities/developments/traction gripper.html,
[Accessed: Jul. 1, 2020].
[39] Fraunhofer, “Flexible single-piece roll-on-gripper for intralogistics,”
[Online] Available: https://www.ipa.fraunhofer.de/content/dam/ipa/de/
documents/Projekte/Forschungsprojekte/300 371 d e.pdf, [Accessed:
Jul. 1, 2020].
[40] J. S. Andersen and T. Christensen, “Apparatus for handling layers of
palletized goods,” October 12 2004, uS Patent 6,802,688.
[41] Wynright Corporation, “Robotic solutions,” [Online] Available: https:
//robotics.wynright.com, [Accessed: Jul. 1, 2020].
[42] M. Bonilla et al., “Grasping with soft hands,” in IEEE-RAS Interna-
tional Conference on Humanoid Robots (Humanoids). IEEE, 2014,
pp. 581–587.
[43] T. Yoshikawa, “Manipulability of robotic mechanisms,” The interna-
tional journal of Robotics Research, vol. 4, no. 2, pp. 3–9, 1985.
[44] T. Yoshikawa, Foundations of robotics: analysis and control. MIT
press, 1990.
[45] C. Della Santina et al., “The quest for natural machine motion:
An open platform to fast-prototyping articulated soft robots,” IEEE
Robotics & Automation Magazine, 2017.
[46] V. Tincani et al., “Velvet fingers: A dexterous gripper with active
surfaces,” in 2012 IEEE/RSJ International Conference on Intelligent
Robots and Systems. IEEE, 2012, pp. 1257–1263.
[47] V. Tincani et al., “Implementation and control of the velvet fingers:
a dexterous gripper with active surfaces,” in 2013 IEEE International
Conference on Robotics and Automation. IEEE, 2013, pp. 2744–2750.
[48] R. Krug et al., “Velvet fingers: Grasp planning and execution for an
underactuated gripper with active surfaces,” in 2014 IEEE Interna-
tional Conference on Robotics and Automation (ICRA). IEEE, 2014,
pp. 3669–3675.
[49] F. Flacco and A. De Luca, “Unilateral constraints in the reverse pri-
ority redundancy resolution method,” in 2015 IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS). IEEE, 2015,
pp. 2564–2571.
[50] A. Palleschi et al., “Time-optimal path tracking for jerk controlled
robots,” IEEE Robotics and Automation Letters, vol. 4, no. 4, pp.
3932–3939, 2019.