IEEE Robotics & Automation Magazine - June 2023 - 49

copy of the object surface, while the Bézier surface fit method
uses a reduced set of design parameters. Furthermore, the latter
features more organic shape characteristics compared to the
projected voxelization-based method, enabling a more robust
grasping performance in the context of position offsets. Both
approaches can potentially be used to automatically generate
or postprocess a training dataset for a machine learning-based
design method, which itself could substitute the slower Bézier
surface fit optimization. This combination of properties makes
the described design approaches suitable for our automatic
design, production, and application pipeline.
PROJECTED SURFACE REPRESENTATION
Similar to [5], the first approach mimics the projected surface
of the desired object via a volume voxelization and a corresponding
point-cloud averaging (Figure 3). (Notice that the
number of points for the object point clouds as well as the
number of visualized voxels have been reduced and partially
varies from picture to picture to improve clearance of the figure.)
This results in a regularly distributed point cloud. To
achieve this, the following steps are conducted:
1) First, the already existing and predefined geometrical data
(STL files) of the desired manipulation object, as well as the
finger base, are imported as point clouds, represented by red
and blue points, respectively, in Figure 3, Base process 1.
The geometrical data represent only the part of the object
that can get in contact with the gripper finger. Note that the
Design Unit
1
3
Production Unit
4
parts inserted into the slots of the IoT-box are correspondingly
not visible and have been cut out in advance.
Additionally, these geometrical data have been remeshed to
increase the level of detail of the resulting point cloud and
the later voxel discretization.
2) After this step, the manipulation object and the finger
base are reoriented and positioned relative to each other
(Figure 3, Base process 2). The objects are aligned to the
middle axis and bottom edge of the finger base. The coordinate
frames of the object points as well as their initial
positions can differ from object to object. The orientation
and position depend on the chosen coordinates/planes
during the CAD design of the objects. The object position
relative to the gripper fingers during the grasp operation
can vary as well. The height between the object's bottom
and the gripper-finger bottom especially can be very different.
Accordingly, the reorientation angles as well as the
height difference between the object and finger base must
be defined manually before the design process. This can
be considered as a start parameter configuration of our
automatic design pipeline.
3) Once the object and finger base are positioned relative to
each other, a designated design space (green box in the figure)
is automatically generated for the fingertip based on a
customizable standard margin between the object and design
space border (Figure 3, Base process 3). In case the object is
higher than the maximum design space, the upper margin is
Task Execution Unit
2
6
A
B
5
B
B
A
A
FIGURE 2. Presented pipeline approach compiled by a design, production, and task execution/evaluation unit. The design unit (1) derives
a fingertip geometry based on a given manipulation object. The production unit produces the final fingers required to conduct the desired
manipulation task based on these generated fingertips. The production is executed by 3D printing the fingertips on preproduced standard
finger base templates (3). These finger bases are stored in designated magazines (2). The extraction from these magazines, as well as
the loading into the corresponding 3D printer finger holder (4) and transfer to the quick-exchange magazines (5), is conducted by robot
arm A. Robot B can then receive and use the fingers stored in these magazines via a quick finger-exchange mechanism. The IoT-box and
grasp-stability test setup (6) enable the application and evaluation of the automatically designed and produced fingers based on three
different target objects: key, battery, and ethernet cable.
JUNE 2023 IEEE ROBOTICS & AUTOMATION MAGAZINE
49

IEEE Robotics & Automation Magazine - June 2023

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