IEEE Robotics & Automation Magazine - June 2023 - 52

serves as an input for the neural network, while the resulting
Bézier surface control points are used as the desired reference
output (Figure 3).
Based on these input and output data, a feed-forward neural
network with two hidden layers (5, 2) using hyperbolic tangent
activation functions is trained.
To generate the Bézier fingertip design for the battery test
object by our neural network, we conducted the described base
process to generate the corresponding input for the network.
The surface generated by the network was then compared with
the Bézier surface fit optimized version using the Hausdorff
distance metric.
EXPERIMENTS
To evaluate our approach, we automatically designed fingertips
for the IoT-box objects of our task execution unit: key, battery,
and ethernet cable (see [18] for IoT-box details). We generated
two fingertip versions for each object by applying the two
described design methods (A: projected surface representation
and B: Bézier surface representation). A series of experiments
was conducted to evaluate the grasping and manipulation performance
of the automatically generated and produced gripperfinger
versions. These experiments are similar to the ones
conducted in our previous work [17]. In particular, one manipulation
series of the experiments covers the following steps:
1) Receive a finger-pair from the quick finger-exchange
mechanism.
2) Conduct the " regular " task at the IoT-box. In this step, the
object is picked up from its storage and inserted into a designated
slot. The IoT-box enables grasp and insertion tests
for three objects: a key, a battery, and an ethernet cable.
3) Conduct the " nonregular " tasks. This step means the same
finger-pair is used to conduct the other (foreign) tasks. For
instance, the fingers optimized for the key will be used to
execute the battery pick and insertion task. These experiments
give insights into the generalization abilities of the
derived fingertips.
4) Conduct a grasp-stability test. Here, the target object is
picked and pushed against a ring structure in two different
directions (Figure 5). If the end-effector crosses the border
of the ring during the push execution, the grasp is considered
unstable.
5) Conduct the regular tasks with pose offset errors for the
corresponding grasping approach pose (Figure 5). In this
step, the target objects are grasped with position offsets
varying from 1 to 5 mm and rotation offsets varying from
Design A
Design B
A B
z
x
y
(b)
z-Offset
(c)
(d1)
(d2)
(d)
(d3)
(e)
(a)
(f)
FIGURE 5. (a) Generated fingertip designs A (projected surface representation) and B (Bézier surface fitting) for the target object's
ethernet cable, key, and battery. (b) IoT-box with manipulation task highlighting pick and insert tasks (pick from designated storage and
insert into target plug, turn in key case). The coordinate frames specify the offset translation and rotation directions used in the posed
offset experiments. (c) Visualization of the z position offset for the key manipulation object. (d1) to (d3) The grasp-stability test setup. The
target object is pushed against the ring surface in two different directions and potential end-effector movements crossing the ring border
are monitored during this operation. (e) and (f) Task execution steps for manipulation object key (e) and battery (f).
52 IEEE ROBOTICS & AUTOMATION MAGAZINE JUNE 2023
Battery
Key
RJ45
IEEE Robotics & Automation Magazine - June 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2023
