Modeling of Actuation Force, Pressure and Contraction of Fluidic Muscles Based on Machine Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper covers the modelling of the force, pressure and contraction of pneumatic muscles. The manuscript collects the dataset based on the information provided by the manufacturer. This data is then used to model using different machine learning techniques. The manuscript is novel - while there are some manuscripts covering the modeling of these type of devices, the authors here focus on different physical quantities, as well as attempting to create a model across different models of the devices. Generally manuscript is well written, but there are some minor concerns I believe need to be addressed.
1. In table 1 - why are there different numbers of points in different datasets? Shouldn't they be similar? Does this affect the scores in a significant way?
2. Some tables are centered and some aren't, this should be corrected.
3. Figures 3 and 4 - how does this distribution affect the models and scores? Please comment on this.
4. There is some small overlap in figures which should be addressed, when looking at main result figures (5 and onward)
5. Conclusion should restate the best results, instead of just stating it's possible to achieve them.
As above are not significant methodological errors, I suggest minor revisions.

Comments on the Quality of English Language

Minor editing is needed

Author Response

We would like to thank the first reviewer on their review of our manuscript. Please find the point-by-point replies to the comments below. Within the manuscript, all the changes have been marked with red text.

 

1. In table 1 - why are there different numbers of points in different datasets? Shouldn't they be similar? Does this affect the scores in a significant way?

 

The following explanation was added to the text:

 

“The differing amount of the points in the dataset stems from the WebPlotDigitizer methodology and the fact that not all dataset images which were used for data collection using aforementioned software have the same amount of data lines present. While the difference in data points can have some influence on the models and scores, all the datasets have a large enough amount of data points as to where it should not adversely affect models. In other words, all datasets are large enough to train the models.”

 

2. Some tables are centered and some aren't, this should be corrected.

 

All of the tables in the manuscript have been centered.

 

3. Figures 3 and 4 - how does this distribution affect the models and scores? Please comment on this.

 

The following text was added in order to address this question:

 

“The histograms show that more data is present in the lower end of the data range. While this is less expressed for contraction, it is very apparent for the combined force dataset (Figure \ref{fig:dist_f}). This may adversely affect the precision of the created models at the higher end of the data range, due to the lower amount of data, and it is safe to assume that most errors will result from this area of data being predicted.”

 

4. There is some small overlap in figures which should be addressed, when looking at main result figures (5 and onward)

 

Thank you for noticing. The overlap between subfigures has been addressed and corrected.

5. Conclusion should restate the best results, instead of just stating it's possible to achieve them.

 

The conclusion was expanded with the following text:

“The overall best results for the combined dataset, meaning for fluidic muscles in general (concerning the fluidic muscles used in the dataset) have been achieved with XGB algorithm for contraction and pressure, achieving mean $R^2$ of 0.90 and 0.92. The best results for force prediction was achieved with MLP, with a $R^2$ score of 0.91.”

English comments: Minor editing is needed

Editing has been performed using spellchecking tools, with minor corrections made throughout the manuscript.

Thank you for your assistance.
Kind regards,

Authors

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors,

Thank you for your interesting paper.

As I can observe, the manuscript includes a discussion regarding the development of a generalized model for fluidic muscles, which is an important contribution towards the development of the subject. For example:

 - The study reflects a clear research gap in the modeling of fluidic muscles by trying to provide a generic model, which might be useful in the industrial fields. 

 - The methods are scientific and use multiple algorithms and validation measures for the sake of making the research design as solid as possible. 

 - Data is straightforward and with proper statistical tests used to explain the results as well as the conclusions made are sound. 

 

However, there are some aspects to which additional work has the potential to make the content clearer and stronger:

 - Some more background information has to be provided in the introduction part, as well as more references to previous research studies. 

 - The methods section is sufficient but can be made more elaborate with detailed descriptions of certain algorithms and validation methods that have been employed for analysis. 

 - There is need to make slight improvement on grammars and make the paper more clear for easy read. 

 

Summarizing, as suggestions I would say: 

 - Enrich the introduction section by providing more details about the context of the study and the associated literature. 

 - Give more information on the methodologies, especially on the algorithms and the methods used to validate these methodologies, so that others can understand and use them again. 

 - The general editing of the manuscript should ensure that those appropriate concerns related to grammar and clarity of the material presented in the manuscript are corrected. 

 

 In any case, the paper is well placed in the literature and, given the identified opportunities, it presents a strong contribution to the field that could be further strengthened.

Thanks again and kindest regards

Comments on the Quality of English Language

English in the paper is overall satisfactory and the ideas have been presented in a systemized and comprehensible manner. Technical vocabulary is employed correctly and linguistic presentation of the paper is helpful to perceive complexity of the techniques and findings.

However, there are several places in which small modifications would make the text more clear and easy to comprehend. Some of the lines can be revised for smoother reading and free from fragmented constructions; there are also some lapses of grammars which should best be rectified to avoid sounding sloppy.

Also, authors should check for a proper use of terms and abbreviations throughout the given body of text since there are differences in some particular cases.

Nevertheless, the language quality does not avoid the flow of the paper and the reader can easily follow the results and the contribution of the paper. Hence, a few minor changes need to be made in order to refine the manuscript to its final touches.

Author Response

We would like to thank the reviewer with their assistance in the review of our manuscript. We believe the comments have helped enrich the manuscript. Please find point-by-point replies below, with the changes in the paper made due to these comments rendered in blue.

1. Enrich the introduction section by providing more details about the context of the study and the associated literature. 

 

We have expanded the introduction in order to provide greater context for the goals presented in the study, focussing on the applications where the proposed algorithms could be beneficial, as follows:

“A highly precise model that can be applied regardless of fluidic muscle dimensions (i.e. one that is developed generally, not for a specific dimension/model of the actuator) can have wide applications. As noted by Hamon et al. \cite{hamon2023model} such muscles can be used as the main actuating components of robotic grippers. Having a rapid controller software that determines the necessary pressure for the given desired force and contraction percentage, would allow for a simpler implementation and the possibility of using a single controller algorithm for multiple dimensions of grippers, greatly simplifying the production and implementation of such end-effectors. Pietrala et al. \cite{pietrala2024design} demonstrate the possibility of usage of pneumatic fluidic muscles as actuators in a 6-DOF parallel manipulator. Authors propose a complex mathematical model for determining the pressure of a muscle necessary to achieve a given position and force. Not only could this potentially be simplified with the application of machine learning, it would simplify the possibility of using fluidic muscles as actuators in serial robotic manipulators, where different dimensions, contraction lengths and forces may be necessary to achieve a desired configuration. Outside of industrial applications, Tsai and Chiang \cite{tsai2023lower}, demonstrate the application of the fluidic muscles for rehabilitation of lower limbs. With differing lengths of limbs in patients, a more general model that will allow a precise determination of the pressure necessary can simplify application of such techniques in rehabilitation. Another possible application of such generalized and high-precision models is wearable tech, as commented by Wang et al. \cite{wang2023novel}. Small high-contraction fluidic muscles can benefit from a more general application by having a single controller for various, differently sized, actuators, allowing for simpler implementation. Final notable application of fluidic muscles that can benefit from a general high-precision model with quick calculation times that could be obtainable from ML algorithms is the antagonistic arrangement, such as noted by Tuleja et al. \cite{tuleja2023analysis}, in which multiple muscles of different sizes are working at the same time in order to achieve a certain configuration. Again, using a single algorithm to control all of the muscles involved in this antagonistic arrangement could simplify the control procedure.”

2. Give more information on the methodologies, especially on the algorithms and the methods used to validate these methodologies, so that others can understand and use them again. 

We have attempted to address this by adding the following text to the start of the section 2.2. Regression Techniques Applied.

“The manuscript applies four algorithms - multilayer perceptron (MLP), support vector regressor (SVR), Elastic Net (ENet) and Extreme Gradient Boosted Machines (XGB). MLP was selected due to its known good performance in various tasks. The algorithms in question were selected due to their known good performance on similar tasks - with MLP being known to provide extremely precise models when a large amount of data is available. In turn, SVR bases itself on the key points in data to create the model, hence eliminating the influence of larger data amounts in certain ranges of data (a possible concern, as will be shown later in the paper), with ENet being selected for a similar reason and its good expected performance on the linear data. Finally, XGB was selected both due to its good performance, but also the ability to provide a set of interpretable tree-shaped models, which may be easier to implement than models resulting from other algorithms.

 

The overview of the methodology is given as follows, with a more detailed descriptions given in the continuation of the paper:

1. 1. the dataset is selected by selecting:

• • the dataset for a single fluidic muscle or the combined dataset,

• • the target value between force, pressure, and contraction (with the remaining two used as outputs)

 

1. 1. the selected dataset is split into five parts for cross-validation procedure,
   2. the loop is repeated as follows:

        (a) out of all possible given hyperparameter combinations one is selected,

        (b) a dataset fold is formed using four parts as training and one as testing set,

        (c) the training set is split into training-validation and the model is created,

        (d) the trained model is evaluated on the testing set with two metrics

 

1. 1. after all the fold have been evaluated the mean scores and standard deviations are noted.
   2. best model hyperparameters are selected based on those scores.

 

 

This procedure is repeated for each of the datasets, as collected in this study. The results are presented in the section 3.”

 

 

 

3. The general editing of the manuscript should ensure that those appropriate concerns related to grammar and clarity of the material presented in the manuscript are corrected. 

As this comment significantly overlaps the English editing comments provided by the author, we have decided to address them together, with the list of changes given below.

 

English comments

 

As noted, some of the abbreviations were not consistent. To address this we performed the following changes:

SVM (Support Vector Machine) was changed to Support Vector Regressor (SVR) in the entire manuscript.

Elastic Net abbreviations was made ENet throughout the paper.

Gradient Boosted Machines abbreviation was rendered as XGB throughout the paper.

The term “pneumatic” muscle was rendered as “fluidic” muscle for consistency.

The paragraph providing the state-of-the-art scores in Introduction was rewritten as:

“Using soft actuators like fluidic muscles requires highly accurate models to predict their behavior. Antonson (2023) \cite{antonsson2023dynamic} compares two modeling techniques, Maxwell-Slip and Bouc-Wen, for predicting muscle force, reporting normalized mean absolute percentage errors of 5.37\% and 12.84\%, respectively. Garbulinski et al. (2021) \cite{garbulinski2021characterization} employed several models—including linear, polynomial, exponential-decay, and exponent—to predict fluidic muscle strain. The respective errors for these models were 4.12\% (linear), 2.10\% (polynomial), 1.94\% (exponential decay), and 1.92\% (exponent). Trojanova et al. (2022) \cite{trojanova2022evaluation} assessed parsimonious machine-learning models for static fluidic muscle modeling. The best models, based on MLP, had errors ranging from 10.10\% to 1.26\%, though each muscle was modeled individually.”

 

Multiple sentences we thought may be unclear were similarly rewritten in the Methodology section, as well as conclusion

In addition to the above, various minor grammar and spelling errors have been corrected throughout the manuscript.

 

Thank you for your assistance.
Kind regards,

Authors