IEEE Robotics & Automation Magazine - June 2023 - 55

show similar performance compared to the computationally
more expansive Bézier surface fit optimization. Therefore, the
learning process can be conducted up front, enabling a comparable
fast design process, while the slower Bézier optimization
must be conducted each time anew.
However, a potential disadvantage of the neural networkbased
design approach is the computation effort to train a specific
network for desired fingertip generation. Such a machine learning-based
method requires potentially large datasets for the
training, of which collection and preprocessing can potentially
consume a lot of time. This can be particularly important if
severe requirements regarding the network generalization abilities
need to be respected. As with all machine learning-based
methods, the generalization characteristics and corresponding
model performance cannot be guaranteed to be always optimal,
due to the stochastic character of the approach.
CONCLUSION AND FUTURE WORK
In this work, we introduced an automatic finger design, production,
and evaluation pipeline. The proposed design
framework represents a critical automation component of
the general production adaption approach, since it further
reduces the time from receiving a new manipulation task to
the ability to manipulate the related object within this
assembly task context. In addition, it decouples the fingertip
design quality from the skill, knowledge, and experience of
the designer. This decoupling, in combination with the
advanced automation level, enables improved scalability of
production setups to different assembly cells and manufacturing
locations since no specialized fingertip designer or
worker must be assigned to each production cell. Overall,
0.2
0.4
0.6
0.8
1
our pipeline constitutes a step further to a fully automated
one-batch size production.
Furthermore, the presented design methods could potentially
simplify and speed up automatic training-data generation
and preprocessing for machine learning-based design
approaches. Our neural network-based fingertip Bézier surface
generation for the battery object illustrates this conceptual
potential. This opens the possibility for an exhaustive
dataset generation and a full comparison of different machine
learning methods, potentially combinable with the described
design methods as well as an experimental evaluation and
comparison in the future.
Remaining challenges of the presented work are the limitation
to simple, nonmechanism-like manipulation objects, the
significant dependency of the assembly line adaptation speed
on the 3D printing time, as well as the manual manipulation
scene (manual object and finger position definition), robot trajectory,
and control parameterization. Depending on the object
size and required grasping paradigm, the form-closure-based
loading and fixation principle potentially limits the adaption
versatility of the setup. This concerns the finger-base holder
on the 3D printer build plate, the quick finger-exchange magazines,
and the loading mechanism design, which enables the
holding of only one finger-base shape and size. Furthermore,
the operational robustness of these loading and fixation structures
depends significantly on the manufacturing quality of
their elements. This in its turn increases the manufacturing
effort of the automatic production setup itself.
Accordingly, future work will focus on addressing these
limitations, as well as improving the overall level of automation
and adaptation flexibility.
Success Rate
0.2
0.4
0.6
0.8
1
Manual Design: Key
Auto Design Method B: Battery Manual Design: Ethernet Cable
Auto Design Method A: Key Manual Design: Battery
FIGURE 9. Experimental result comparison of manually designed fingertips of [17] and automatically generated fingertips of the
presented work.
JUNE 2023 IEEE ROBOTICS & AUTOMATION MAGAZINE
55
key
battery
rj45
grasp_stability_t...
pos_y_4
pos_x_4
pos_z_5
pos_y_5
pos_x_5
rot_z_5
rot_y_5
rot_x_5
rot_z_10
rot_y_10
rot_x_10
rot_z_15
rot_y_15
rot_x_15
pos_z_1
pos_y_1
pos_x_1
pos_z_2
pos_y_2
pos_x_2
pos_z_3
pos_y_3
pos_x_3
pos_z_4

IEEE Robotics & Automation Magazine - June 2023

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