Page 72 - IJAMD-1-1

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

was designed to ascertain specific output parameters, 2.3. Numerical modeling and simulation of gripper
specifically the angular displacement that resulted from 2.3.1. Modeling of gripper
variations in the aforementioned input parameters.
In the original design, as indicated in Figure 3A, the finger
Subsequent to the simulation phase, we constructed comprised intricate features such as fins and cleats along
a comprehensive dataset derived from the simulation its length, with interconnects that likely facilitated the
results. This dataset encompasses a range of data points integration of sensors or actuation mechanisms. While
that realistically represent the operational spectrum of the these features could potentially enhance performance,
gripper, allowing us to illustrate the correlation between they added complexity to the simulation models. We
input variables (pull distance and joint temperature) and simplified the numerical model for the simulation analysis
resulting angular displacement outputs. by removing non-critical features, ensuring it accurately
With this dataset in place, we proceeded to the machine represents the gripper finger’s essential joint bending
learning stage, where we applied a suite of algorithms, behavior without unnecessary complexity.
namely linear regression (LR), decision tree regression Figure 3B shows the simplified model used for the
(DTR), and k-nearest neighbor regression (KNN), to simulation, with the fins, interconnects, and cleats
discern patterns and relationships within the data. The removed from the structure. This simplification likely leads
choice of these particular machine learning models was to a reduction in the computational resources required for
strategic, as each presents different strengths in terms of simulation and modeling, allowing for a more efficient
prediction accuracy, interpretability, and computational dataset development process.
efficiency.
The ultimate goal of this methodology was to develop 2.3.2. Simulation of gripper
an accurate predictive model for actuation control. Using To simplify the material characterization process, isotropic
angular displacement as a primary input, this model linear elastic material properties were assumed for all
predicts the necessary pull distances and joint temperatures, parts of the model, which were obtained from uniaxial
effectively optimizing the gripper’s performance. This tensile tests and are presented in Table 1. In this study, we
capability allows tailored movements for the task at hand approximated Young’s modulus of conductive polylactic
without relying on computationally expensive simulations. acid (cPLA) across a range of temperatures using the
Through this predictive model, we aim to achieve a level temperature-dependent storage modulus derived from
of control and adaptability that enhances the functional dynamic mechanical analysis (DMA) of the cPLA material.
capabilities of the soft robotic gripper. Dynamic mechanical analysis was conducted on the cPLA

A

B

Figure 3. A comparison between (A) the original gripper design and (B) the simplified model used for numerical simulation.
Note: The images are for illustration purposes only and are not to scale.
Abbreviations: cPLA: Conductive polylactic acid; TPU: Thermoplastic polyurethane.

Volume 1 Issue 1 (2024) 66 https://doi.org/10.36922/ijamd.2328