IEEE Robotics & Automation Magazine - September 2020 - 70

was the case with the carrot in Figure 15, which was grasped
with only two fingers instead of four. Finally, in the accompanying video on IEEE Xplore, one can see that the SoftHandler end
effector adapts to the apple position.
General Discussion
The "System Characterization" section compared the SoftHandler with a comparable rigid manipulator, and the "System Validation" section presented a validation of the system
when grasping objects that presented a broad variety of physical and geometrical characteristics. During the experiments,
we observed that some of the grasp may fail during the execution of the task. Most of the unsuccessful grasps observed
were due to two types of failure: slipping and missing.
We refer to slipping when the grasped item slips out of the
gripper fingers. This issue occurred, for example, in experiment 3 (Figure 15). The failure is linked to grasps for which
the fingers do not completely cage the object. This may happen when the displacement between the end effector and the
object is large. As suggested by Figure 8(i) and (h), tiny
objects and heavy objects are the most challenging, especially
when the displacement is large. Another reason for slipping
can be a wrong grasping approach. In all of the experiments,
we employed the same grasping approach without any specific tuning for the item to be picked. Future work will analyze
different grasping approaches for varying object geometries.
On the other hand, we refer to missing when the end effector is placed where there are no objects. This issue is related to
the vision system passing to the control system an incorrect
position of the fruit to be picked. If we ignore this type of fault,
the success rate increases from 83 to 100% [experiment 1, Figure 14(a)], from 78 to 100% [experiment 2, Figure 14(b)],
from 69 to 81% (experiment 3, Figure 15), and from 87 to
100% [experiment 4, Figure 16(a)]. During experiment 5 [Figure 16(b)], no vision errors occurred. Thus, the average grasping success rate, neglecting vision errors, was 96%.

Interestingly, this result is close to the success rate
obtained in the "Grasping Performance" section. Indeed,
the average positioning error of the end effector during
experiments 1-5 is 23.58 mm, which corresponds to a
success rate of 99%, according to Figure 8(l). The difference between the two values can be ascribed to experimental variability and differences in the setup, which
include different controllers and manipulated objects.
Another issue to consider is the double-grasp, that is,
when the gripper closes around more than one object. This
tends to occur when the objects are cluttered (e.g., Figure 15),
and the adopted vision system merges close items into one
cluster. A double-grasp may result in either no picking, picking a single item, or picking more than one. Each of these
results could be considered a failure or not, depending on the
task. Employing a different grasping approach or a different
vision system may solve this behavior.
The quantitative experiments in the "System Validation"
section let us appraise the average picking time of the SoftHandler, which amounts to approximately 10 s. The most
important factor limiting the speed of the system is the
power of the adopted actuators, which are mainly intended
for research purposes. Figure 4(d) shows that their maximum velocity is lower than the that of most industrial
manipulators. On the other hand, our goal was not to develop a system highly optimized in terms of speed performance but rather to explore the development of a device for
handling heterogeneous and fragile objects that is robust to
accidental impacts. Future work will aim to improve the
actuation and, more generally, system performances.
A second factor limiting speed is linked to the vision system. Each failure in the vision system leads to an additional
cycle for emptying the crate, resulting in a longer task completion time. The experiments in the "System Validation" section
showed that the employed vision system presents an average
failing rate of 15%.

Figure 15. A photo sequence of the raw-food handling scenario in experiment 3. The SoftHandler empties a crate full of toy fruits.
The objects differ in shape and size, and they are randomly placed within the crate. The size range of the objects is Q25-80 × 50-170 mm,
while the weight range is 8-20 g.

70

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

SEPTEMBER 2020
IEEE Robotics & Automation Magazine - September 2020

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