Page 69 - IJAMD-1-1

International Journal of AI
for Material and Design Machine learning for gripper state prediction

such as pneumatic inflation. , shape memory alloys. , present unique challenges in modeling due to the non-
4,5
2,3
and electroactive polymers. Pneumatically actuated soft linear and temperature-dependent material properties,
6,7
grippers benefit from simple, fluidic control and can exert such as those exhibited by polymers like thermoplastic
gentle, tunable gripping forces, rendering them suitable polyurethane (TPU) and conductive polylactic acid
for handling fragile objects. However, their practicality is (c-PLA). However, the limited understanding of the
hindered by the need for complex tubing and limitations interplay between gripper geometries and design, material
imposed by the constant need for air supply and low properties, environmental factors, and actuation control
gripping force, making them less ideal for mobile or poses a significant challenge in the development of reliable
heavy-duty applications. In contrast, electroactive prediction models. Moreover, customized gripper designs
3,8
polymers and shape memory alloys offer precise control can be achieved with 3D printing, adding challenges to
and can be activated electrically, which simplifies the the development of a systematic approach to predict the
system’s integration with control circuits. Nevertheless, grippers’ unique dynamics and kinematics. Although
22
these materials may pose challenges due to potential estimation of the dynamics and kinematics of grippers
costliness and limitations in providing the necessary force with variable stiffness has been successfully demonstrated
or durability for specific applications. 6 through a predictive model developed based on simulation,
Cable-driven soft grippers emerge as a robust alternative it is limited to systems with decoupled actuation where the
within this landscape. Their strength lies in the mechanical dynamics of one joint are independent of the condition
9
21
simplicity and direct control offered by the cable actuation and state of the other joints. For gripper systems with
mechanism, which allows for precise manipulation of the more complex dynamics, the gripper’s state estimation or
gripping motion. 10,11 This approach provides a high level modeling often relies on either a trial-and-error approach
of dexterity and can be easily scaled from micro to macro or exhaustive empirical methods that are impractical for
applications. Moreover, cable-driven systems bypass real-time applications.
12
the need for onboard compressors or complex electrical Machine learning is an important tool in modern
drivers required by other soft actuation methods, favoring technology, used to make predictions by analyzing data.
23
a straightforward design that is both lightweight and agile. This technology is widely used for various purposes, such as
The cable-driven design also permits the integration of predicting the quality of manufactured parts. or reducing
24
variable stiffness features, further enhancing the gripper’s reliance on human intervention to improve detection
adaptability. This adaptability enables it to conform to or inspection accuracy. In the context of robotics,
13
25
and securely grip objects of various shapes and sizes. specifically in soft robotics, the application of machine
Combined with the potential for intricate feedback loops learning is transformative. It enables the development
26
and advanced control algorithms, cable-driven soft of predictive models capable of anticipating the behavior
grippers position themselves at the forefront of the field, of systems characterized by a high degree of complexity
particularly in situations requiring nuanced or sensitive and variability, such as soft robotic mechanisms. 27,28
interactions with the environment. 14 The cable-driven soft gripper, as discussed in this paper,
Gripper state prediction is crucial for enabling precise exemplifies such a system. Machine learning algorithms
control and dexterity in robotic manipulation, directly effectively navigate the intricate dynamics of these devices,
impacting the efficiency and adaptability of automated considering the non-linear properties of the materials and
systems in complex, real-world tasks. The current state- their multifaceted interaction with external stimuli.
15
of-the-art gripper state prediction has primarily focused Addressing the challenges mentioned above, this
on rigid systems, where standard robotic kinematics paper explores the development of a predictive model
16
and dynamics can be applied. However, soft grippers, for precise control of a cable-driven soft gripper with
such as bionic finger grippers, challenge these traditional variable stiffness joints using machine learning techniques.
models due to their inherent compliance and infinite Employing advanced finite element analysis with ANSYS
degrees of freedom. 17,18 This challenge has necessitated alongside machine learning techniques, we have developed
the development of innovative approaches to predict and a comprehensive framework for analyzing and predicting
control their behavior. the bending angles of gripper joints by utilizing a novel
In light of these complexities, this paper advances the methodology that synergizes data-driven models with
research frontier by addressing the critical need for accurate physical simulations. We conducted a comparative analysis
bending angle prediction in variable stiffness grippers based of different regression models to predict the joint bending
on shape memory polymer. 13,19-21 These devices, operating angle of a gripper, factoring in variations in joint stiffness
fundamentally differently from their rigid counterparts, and the corresponding effects at varying pull distances. Our

Volume 1 Issue 1 (2024) 63 https://doi.org/10.36922/ijamd.2328