Page 78 - IJAMD-1-1

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

Table 4. Joint angles simulated using the joint temperatures and pull distance predicted by the linear regression, decision tree
regression, and k‑nearest neighbors regression models. The highest absolute error and mean absolute error are based on the joint
angles listed in Table 3

Case Linear regression Decision tree regression K‑nearest neighbors regression
Joint 1 angle (°) Joint 2 angle (°) Joint 1 angle (°) Joint 2 angle (°) Joint 1 angle (°) Joint 2 angle (°)
1 17.6* 20.2 11.9 23.7 8.1 31.9
2 16.0 17.0 17.7 17.1 19.1* 17.0
3 35.9 39.3 32.2* 35.5 35.2 42.0
4 34.3 40.5 41.6 55.7 29.7 50.8
5 44.1 51.3 52.3 60.0 50.3 60.0
Highest absolute error* (%) 75.7 23.2 27.1
Mean absolute error (%) 20.6 14.7 10.8
Note: *The case with the highest absolute error for each model.

explained by three lines of equations. Therefore, DTR is nor did they account for potential thermal annealing
not the preferred choice among the three algorithms, as it effects on the cPLA material, which could lead to an
offers fewer improvements and does not provide further increase in stiffness over time. In future research endeavors,
insight into the relationship between the variables. On a efforts are anticipated to obtain the experimental data for
side note, while KNN demonstrated the best performance, Young’s modulus of the materials, expand the dataset,
it is the most challenging to interpret, as visualizing KNN refine the machine learning parameters, and explore the
becomes challenging when the dimension of the dataset is incorporation of the thermal annealing effect into the
higher than 2. dataset for more accurate data representation. In addition,
The performance of the fitted models was validated there is an imperative need to further validate the models
using five new untouched data points. First, the models using physical prototypes to ensure the practical viability
predicted the temperatures of joint 1, joint 2, and the pull of the predictive algorithms in real-world scenarios.
distance based on the required angles as input. The resultant 4. Conclusion
angles of the two joints were obtained from simulation
using these parameters, which were then compared to the This study presents insightful revelations into the dynamics
required angles. Table 3 shows the predictions provided by of a stiffness-tunable gripper, particularly focusing on the
the three models for the required joint temperatures and correlation between joint temperature (and consequently
pull distance to achieve the designated joint angles for each joint stiffness) and bending angles at various pull
case. The comparison of the performance among the three distances. Through simulation and experimentation, a
2
models is shown in Figure 8D. All three models exhibit an R comprehensive evaluation of a spectrum of regression
score close to one, verifying the effectiveness of the models models was conducted to identify the most effective ones
in obtaining the parameters to achieve a certain targeted for predicting the behavior of cable-driven soft grippers
angle for both joints. The analysis of the performance of across diverse operational conditions. Our findings
the fitted LR, DTR, and KNN models and the errors in the indicate that the KNN regression model outperforms its
joint angle prediction is summarized in Table 4. The KNN counterparts, offering a superior balance of accuracy and
model outperformed the other two models by having the computational efficiency, maintaining an MAE below 11%.
lowest MAE for both joints, consistent with the results from The proficiency of this model underscores the potential of
the cross-validation, proving that KNN is the best model machine learning in the realm of soft robotics, bridging
among the three approaches. Based on these 10 predicted the gap between theoretical constructs and their tangible
joint angles, it is observed that the highest absolute error applications. The successful prediction of joint bending
was approximately 27.1%, and the MAE was identified to angles is not a mere academic exercise; rather, it represents
be 10.8% for the KNN model. a substantial leap toward the realization of advanced
automation systems. The enhanced precision in soft
3.3. Limitations gripper control paves the way for a new echelon of robotic
This study acknowledges the inherent limitations within applications, particularly in sectors where traditional rigid
the current scope of data and model complexity. The grippers prove inadequate for tasks requiring nuanced and
current models did not use actual Young’s modulus data, delicate handling.

Volume 1 Issue 1 (2024) 72 https://doi.org/10.36922/ijamd.2328