IEEE Robotics & Automation Magazine - June 2023 - 47

products is very laborious and time-consuming. Several steps,
such as changing and producing the robot's end-effector, changing
the toolset, and reprogramming the robot must be conducted
manually. Therefore, manufacturing lines are usually optimized
for mass production of the same components. A small-batch-
size production would require a significantly shorter adaptation
time, ideally in the form of a fully automated process. Accordingly,
the robot arm and end-effector setup must be able to
quickly adapt to a new spectrum of components, as soon as a
new and potentially unknown product line could be assembled.
Parallel grippers are still mainly used as end-effectors for
robot arms with industrial application since potential alternatives
are not feasible yet. For instance, complex multifingered
manipulators are still a subject of research and soft material
manipulators have critical robustness issues when manipulating
heavy objects with sharp edges.
The fingertips of these parallel grippers must be adapted to
the objects that will be grasped and manipulated by the robot
to achieve the targeted performance, robustness, and grasp stability.
This finger adaptation is mostly done manually, using
classical production and fast prototyping methods, as well as
services, such as machining or 3D printing. The manual adaption
and production process has the following disadvantages:
■ The new fingertips performance depends on the designer's
skills, knowledge, and experience [1], which is a problem
from a quality consistency point of view.
■ Designing, manufacturing, and mounting the fingers to the
gripper are very costly in terms of time and labor.
Depending on the object and task complexity, several iterations
are often required to finalize the fingers. During
each iteration, costly steps like designing, manufacturing,
and testing must be conducted over and over again.
■ Due to these manual, nonautomated operations, the
approach poorly scales to an arbitrary number of components,
assembly setups, and production locations.
To reduce finger design time, different guidelines and strategies
have been developed. Greg Causey proposed a list of
design guidelines to maximize the grasp-stability and manipulation
performance of parallel gripper fingers [2]. A spectrum
of approaches that try to automate the gripper-finger design has
been introduced in [3], [4], [5], [6], [7], and [8]. For example, Ha
et al. optimized the finger geometry via a neural network-based
design framework [1]. Song et al. derived the fingertips based
on a Bézier surface representation and fitting [6]. Honarpardaz
et al. used a projection-based design approach to automatically
create gripper fingertips [5]. Velasco and Newman applied subtractive
Boolean CAD operations to realize form-closure-based
automatic gripper-finger design [7].
However, not only the finger adaptation process but also
the finger-gripper system itself limit the robot-production setup
adaptation flexibility and performance. In general, a robot with
a permanently mounted finger-pair can only manipulate objects
for which these fingers are designed. Accordingly, production
line setups are built-up by many robots, which conduct one
single task and manipulate only one single object per robot. To
cope with this issue, several tools- and end-effector changing
systems have been developed [9], [10], [11], [12], [13], [14]. These
approaches focus on changing the entire end-effector automatically,
which potentially increases costs in case a larger number
of objects must be manipulated. Note that systems that enable
single-finger change require manual operations [15], [16].
As a partial solution to the tool-changing and adaptation problem,
we developed a designated gripper-finger production and task
execution setup [17]. This setup is able to produce new fingers by
printing custom fingertips on standardized finger bases quickly
on-demand. The fingers can be stored in prepared magazines and
used by another robot arm as desired. This robot is equipped with
a specially developed quick finger-exchange mechanism, which
enables autonomous mounting and demounting of corresponding
fingers provided within the mentioned magazines. Accordingly,
the same robot can change the used finger-pair during an assembly
process in case the robot must manipulate another object. This
not only decreases the finger adaptation time but also increases
the robot task execution flexibility.
Using our proposed quick-exchange mechanism, a substantially
smaller number of robots is needed to complete an assembly
task since one robot can manipulate many different objects
robustly by using different specialized finger-pairs for each
object separately. Despite these improvements, the setup in [17]
requires manually designed fingertips to successfully grasp and
manipulate the designated objects. As previously mentioned,
this couples the task execution success to the capabilities of the
designer and reduces scalability while increasing time and costs.
Therefore, in this work, our aim is to introduce a framework
that allows a fully automated gripper-finger adaptation process.
To do so, we implemented an automatic fingertip design pipeline
and integrated it into our previously developed automatic
finger production and application setup of [17] (Figure 1). Based
on the existing geometry information of the given manipulation
objects (CAD data), the proposed pipeline is able to automatically
derive the required fingertip design. Leveraging some
results in [5] and [6], we implemented two different main versions
of this pipeline, as well as a separate machine learningbased
design method. The approaches are form-closure-based
and try to mimic the surface of the desired manipulation objects.
The first approach is mirroring the object's surface one by
one, using a voxelization and projection method to copy the
surface negative to the fingertip body, analogical to [5]. The
second approach is fitting a Bézier surface to the discretized
object geometry, similar to [6]. A third approach, which uses
the results of the previous design methods to train a neural network
and substitute the slower Bézier surface fit optimization,
was implemented to illustrate the potential data generation
and processing advantages of the two main design methods
for machine learning-based design approaches.
The combination of these automatic fingertip design approaches
with our previously developed automatic production and application
testbed [17] results in a fully automatic gripper-finger design,
production, and application pipeline with, to our best knowledge,
an unknown level of on-demand production adaptation performance.
The presented pipeline can be considered as a step toward
fully adaptive small- or even one-batch-size production since it
JUNE 2023 IEEE ROBOTICS & AUTOMATION MAGAZINE
47
IEEE Robotics & Automation Magazine - June 2023

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