Page 77 - IJAMD-1-1

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

complex heat transfer dynamics to more accurately reflect
real-world conditions. Pull distance 10.0 10.0 20.0 20.0 30.0
3.2. Prediction of joint angles using machine (mm)
learning techniques
The performance of every machine learning algorithm
obtained from the cross-validation step is shown
in Figure 8A. Unsurprisingly, LR exhibits the worst K‑nearest neighbors regression Joint 2 temperature (°C) 70.0 47.5 62.5 67.5 62.5
performance among the three algorithms, with the
highest MAE for every predicted output. In contrast,
KNN demonstrates the highest accuracy among the
three algorithms. These results further support the earlier
hypothesis that bending behavior is closely related to Joint 1 temperature (°C) 52.5 60.0 62.5 65.0 52.5
the specific configurations of stiffness and pull distance.
Notably, there is a significant improvement in the accuracy
of pull distance prediction when using KNN instead of LR.
The MAE is reduced from 0.71 ± 0.21 mm to 0.17 ± 0.27 mm,
representing a 76% reduction in MAE. In comparison, the Pull distance (mm) 10.0 10.0 20.0 30.0 30.0
MAE for the predicted output of the temperature of joint 1
and joint 2 is reduced by 31% and 35%, respectively, from
7.93 ± 1.55°C to 5.50 ± 1.85°C for joint 1 and 7.94 ± 1.52°C
to 5.13 ± 1.68°C for joint 2. As depicted in Figure 8B, the Joint 2 temperature (°C) 70.0 50.0 65.0 65.0 70.0
Spearman correlation is remarkably high between the Decision tree regression
joint angles and the pull distance compared to the other
two outputs. This observation suggests that KNN is more
effective in obtaining the targeted angle, as evidenced by
a significant improvement in the MAE of pull distance. Table 3. The required joint temperatures and the pull distance predicted by the trained linear regression, decision tree regression, and k‑nearest neighbors regression
Meanwhile, the weak correlation score between the joint’s Joint 1 temperature (°C) 60.0 60.0 65.0 60.0 60.0
temperature and the joint’s angle may be attributed to the
non-linear relationships between these variables.

While the LR model exhibits the highest MAE among
all three algorithms, it offers good explainability for the fit. Pull distance (mm) 10.9 9.5 22.7 22.6 29.6
The fitted model is represented by these equations:

Joint 1 temperature, y =52.95+0.28x -0.21x (II)
0 1 2
Joint 2 temperature, y =54.10-0.35x +0.33x (III) Joint 2 temperature (°C) 59.9 55.2 54.4 58.4 56.2
1 1 2 Linear regression
Pull distance, y =0.44+0.26x +0.26x (IV)
2 1 2
From the equations, it can be inferred that the required
pull distance is longer when both joints are more bent. In
addition, a higher temperature on joint 1 is needed for Joint 1 temperature (°C) 48.8 53.5 56.0 51.7 54.8
joint 1 to bend more than joint 2, and vice versa for the
temperature on joint 2.
While DTR has a lower MAE compared to LR, models to achieve the designated joint angles
interpreting a DTR model is more challenging. As Joint 2 angle (°) 30 20 44 50 60
depicted in Figure 8C, although the MAE of the DTR is
lower with an increased hyperparameter of max depth, Test data points
the model becomes increasingly complicated. An example Joint 1 angle (°) 10 15 42 35 52
of the fitted DTR model is shown in Supplementary file,
and even with just a max depth of five, it is challenging
to interpret the model compared to LR, which can be Case 1 2 3 4 5

Volume 1 Issue 1 (2024) 71 https://doi.org/10.36922/ijamd.2328