IEEE Robotics & Automation Magazine - June 2021 - 68

(a)
(b)
(c)
(d)
(e)
(f)
(g)
Figure 10. Task execution in the first experiment at t = (a) 0, (b) 6.5 (corresponding to the coarse pregrasp pose ()-1
by the object-detection module), (c) 18.5, (d) 21, (e) 28, and (f) 30.5 s. (g) The end of the object-pushing phase.
TTj
b
t
v
j
l
fk+= -+ (11)
nn
a
() () ()
11 fk
a ()
n
fkt
,
where the value α sets the time constant of the filter and () isfkt
the measured tangential force. The interested reader can find
more details about the slipping-control algorithm in [16] and [18].
Experiments
The experimental setup shown in Figure 1, emulating a
supermarket shelf, comprises a KUKA LBR iiwa 7-kg robot
manipulator; a Schunk WSG50 gripper equipped with force/
tactile fingers designed and produced in our laboratory, the
so-called SUNtouch [15]; and an Intel D435i RGB-D camera
mounted on the last link of the robot. The whole robotic system
programming runs under ROS.
Objects P5 and P7 are used in the following experiments,
with the physical parameters reported in Table 1 together
R3
3
2
1
-1
-2
-3
20
22
24
26
Time (s)
Figure 11. The planned and actually executed motion of the robot joints in the first
experiment.
68 * IEEE ROBOTICS & AUTOMATION MAGAZINE * JUNE 2021
28
m31
= 0
R4
m41
= 0
estimated
with the few parameters of the reactive control algorithms: λ
for the visual-servoing controller in (5), α for the pivoting
controller in (11), and l for the observer in (8).
The first experiment shows the task consisting of finding
product P7 on the pick desk (see Figure 9) and grasping and
placing it on the central facing of the top shelf on the right
(Figure 1), as shown in Figure 10 (a), where the robot is in its
initial configuration. The object-detection module is interrogated
with the CNN corresponding to product P7. The
bounding box and coarse orientation, identified by the CNN,
are used to roughly place the gripper close to the pregrasp
pose of that object
^ht .s .65=
Given the coarse orientation,
the corresponding plan is fetched from the database.
The plan result of requests R4
and R5
is shown in
R5
m51
Figure 11 with dashed lines; the previous planned trajectories
are not shown here because they are not used in the execution
phase. As explained in the " Execution Phase " section, the
visual-servoing controller is activated
to bring the gripper from the coarse
pregrasp pose
= 0
pose ^ht 18 5=
.s . At t 18 .5=
^h= 65 to the grasp
s, the
t .s
gripper is asked to grasp the object,
while the arm is asked to lift it a few
centimeters to move away from the
tray, and the grasp controller automatically
computes the grasp force using
the control law in (10).
At about t 19=
s, the robot starts
q (t) (Executed)
q (t) (Planned)
30
32
lifting the object, and a peak of about
-1.6 rad/s in the estimated slipping
velocity is detected by the tactile fingers
(Figure
12). A quick 10-N
increase in the dynamic grasp force fnd
is recorded, which allows the robot to
safely lift the object without any significant
slippage. This is confirmed by the
(rad)
"
IEEE Robotics & Automation Magazine - June 2021

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2021
