IEEE Robotics & Automation Magazine - December 2020 - 21

the soft gripper, which, subject to high acceleration, is not
capable of conforming to the handle and establishing a firm
grip. By using the arm and the legs together, we recorded a
pulling force of 15 N [Figure 7(c)], similar to that achieved
by the arm alone. By activating the arm first, the robot
positioned the gripper in a suitable grasping position, so
the subsequent activation of the legs fully exploited the
potential of the gripper. By activating the legs very slowly
and allowing the soft gripper to conform well to the handle
[Figure 7(d)], we recorded an average force of approximate-
ly 18 N, confirming that slow acceleration enables a greater
pulling force. In this case, the power consumption was
reduced to a maximum value of roughly 96 W, even though
the motion was maintained for 6 s, with an average power
draw of 70 W [Figure 7(c)-(d)].
We also measured the power consumption of the soft
manipulator (while grasping a shell with a size similar to
the one that had been gripped on the seafloor) in the labo-
ratory water tank. The system's power was recorded every
5 s by a power meter device (Zhejiang Shixin Electric). We
calculated the average power consumption of the soft
manipulator by recording data for 40 s. We also repeated
each experimental scenario five times to acquire the mean
value. The pneumatic control system's average power con-
sumption was 25 ± 1 W. This result highlights a less-studied
aspect of using a soft gripper for underwater grasping.
Most existing research examines novel actuation methodol-
ogies, modeling, sensing, and fabrication methods. Our
results indicate that the arm's speed and acceleration greatly
influence the maximum grasping force. In addition to
underwater operations, this finding may be particularly rel-
evant for factories, warehouses, and other industrial set-
tings that require quick handling.
Finally, the pulling force of the legs was measured by
directly connecting a cable between the force sensor and
the robot's body, without the involvement of the soft arm.
Theoretically, the stall torque of the motors (3.8 N∙m)
exerted on a tibia length of approximately 0.35 m should
result in a force of roughly 65 N for all the legs together.
Actual measurements from experiments reported an
average force of 44 N [Figure 7(e)]. The ratio between
the measured and nominal forces is roughly 0.68, con-
sistent with the expected losses from a servo motor sys-
tem. For underwater vehicles, even with optimized blade
designs, the efficiency would be reduced by the propeller
efficiency to approximately 0.8 [29]. For an optimized
AUV, the overall power efficiency for 7-N navigation is
reported as 0.49. It is a reasonable expectation that direct
transmission would slightly increase the force manage-
ment and possibly exert a higher force on the environ-
ment. However, an ROV of comparable size to SILVER2
(the BlueROV2) demonstrates a nominal bollard vertical
thrust of between 69 and 88 N. Without the necessary data
for comparison, we can assume that the force and power
consumption of SILVER2's legs alone fall between those of
a small ROV and an I-AUV.

Workspace Protocol
An intrinsic advantage of mobile manipulation is that loco-
motion enables the manipulator workspace to be extended.
Our protocol investigated vertical extensions first, as shown in
Figure 8. In this case, the robot moved into a starting position
and initiated a test. Compared with the soft manipulator
working alone, we found that the vertical workspace of the
soft manipulator increased 85% (from 20 to 37 cm) when it
was integrated on SILVER2. With an average release error of
6.1 cm, the arm was capable of placing a portion of the shell
in the red target area. Such an error is comparable to remotely
controlled tests performed in the swimming pool [6], where a
yellow tube was grasped during a collection task. In that
example, precise positioning was not required if the mission
was satisfactorily completed.
In previous work [5], simulations were performed for
button-pushing and valve-turning tasks. The reported
alignment error and the gripper and button modeling were
slightly better than our present results. However, our results
suggest that the button-pushing task could be accom-
plished, even with our average placement offset. In [30], the
authors report good results for the end effector positioning
within an experimental tank and under a monodirectional
simulated current. In that experiment, six cameras and
Qualisys Track Manager (QTM) software were employed to
evaluate the system's underwater pose. With an error of
approximately 5.8 ± 3 cm, the experimental results demon-
strate very accurate mobility and positioning. It is worth
noting that the testing conditions presented in our experi-
ments are poorly represented in the existing literature. All
related work either accomplishes the target task through
simulation or in controlled environments, such as a swim-
ming pool or a tank. It is correctly reported in [30] that the
position error via the QTM is difficult to obtain in field
conditions, so the reported performance could vary signifi-
cantly among different underwater environments.
Mobile Pick and Place
Compared to the results of the workspace protocol, the per-
formance decreased in the horizontal workspace extension
during an evaluation known as the pick-and-place test. In all
attempts, the robot succeeded in bringing the target within
the manipulator workspace (i.e., between the legs). During
the release, we observed an average error of roughly 15.7 cm
[Figure 9(d)]. This result enabled us to place the object in the
desired area but not always within the red, circular target. We
observed a primary reason for this decrease in the perfor-
mance: disturbances on the benthic platform and the soft
manipulator. SILVER2 does not have closed-loop stationkeeping control, and we did not modify the position of the
robot when unexpected currents significantly shifted it from
the desired final location. With respect to the arm, although
we never observed significant displacements, small vibrations
may have impaired the release speed and increased the error.
It is worth mentioning that, in one trial, the shell was acci-
dentally released during the walking phase from the starting
DECEMBER 2020

IEEE ROBOTICS & AUTOMATION MAGAZINE

IEEE Robotics & Automation Magazine - December 2020

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2020