Generative Design of Soft Robot Actuators Using ESP

Round 1

Reviewer 1 Report

I would like to congratulate authors for performing the impressive study; however, there are some major improvement which needs to be done before the work is evaluated. 

I have some following suggestions and questions which authors need to follow before the consideration of manuscript. 

Please find the attached document. 

Comments for author File: Comments.docx

Author Response

Thank you for your review. Please see my comments below. I have organised your comments in an enumerated list below. With my comments italicised below each. 

1.  Authors need to improve the Abstract section to add their contribution in this manuscript clearly.
   1. Thank you for the observation. I have updated the abstract and conclusion to highlight the contribution of this paper. The contribution demonstrated is that combining the Encapsulation, Syllabus, and Pandamonium method with a reduced order model or other encoding makes for a practical generative design process that produces results comparable to the existing state of the art, at most the same design time while including the human designer meaningfully in the design process and facilitating the inclusion of other numerical techniques such as MCMC.  Please see the attached marked-up version of the paper highlighting all changes made.
2. Authors need to improve the Introduction and materials section and using these following references which are quite popular and getting popular as well. I recommend authors to reduce the unnecessary number of references and try to find some review papers which can explain the pointwise characterization of soft actuators, their applications as well as materials. There are more than 50 references which is cited in the introduction section only which is not that much necessary. Some of them are mentioned below; however, there are plenty amount of review articles available which explains all of these types of reviews:
   1. My understanding of this point is that the reviewer would prefer that I change the style with which I present prior art by removing the references to specific papers of interest in favour of a shorter list of review papers. My intention with the introduction section is to highlight the breadth of the field with examples that contribute to my chosen application later in the paper. I would also prefer a shorter list of references but still want to credit the original authors rather than the authors of derived work, such as review papers unless referring to some conclusion made in the review not presented in the component papers. Per your suggestion, I have thinned out the long list of references, selected the most important examples for my context, and added detail which refers to some of the references proposed. I have also moved away from the bullet point presentation, with detailed descriptions of each application and challenge, in favour of a conventional paragraph format. Please see the attached marked-up version of the paper highlighting all changes made. 

1. Authors clearly need to improve the citing style which is being cited by journals. There is no need of putting links of the journals in every references.
   1. Per the reviewer's suggestion, the references have been styled more consistently, and all URL links have been removed. Please see the attached marked-up version of the paper highlighting all changes made.
2. I would like to suggest some of the articles which authors need to go through and explain about challenges mentioned in these following articles: Thank you for pointing out that a critical paragraph in my introduction is unreferenced, and thank you for suggesting references for inclusion that support my narrative. I will address these additional reference suggestions in more detail below. Please forgive me for not copying the references in my response but only your comments. 
   1.  In the following article, the authors explain the challenges the opportunities of performing real-time finite element simulations. How do the authors address these challenges or solution in their article:
      1. This is an interesting comment. My research group generally focuses more on materials calibration methods than robotics, and the introduction of Physics informed neural networks is exciting, especially for this class of materials. We are currently looking at how to incorporate this into our research. As yet, it is unclear how this will be done, and as far as I am aware, no commercial FE solver has a stable implementation of a material model trained this way. In my opinion, and I stand under correction, PINs are not yet ready for implementation in an application like this, but they are a very promising technology I look forward to exploring. In response to your comment, I have included a note on material calibration as a major source of error in soft robots simulated using FE tools. I also included some formal details of FE model we used for the paper under the reduced order model section. 
   2. Recently, the few authors focused more on force shape estimation and measurement when it comes to design the force/pressure sensors modelling and then comparison with the sophisticated experiment using the real-time and CAD simulators. \\ In these following articles, author perform real time simulation and experiment of different robotic sensors and phantoms, I would suggest authors to go through these articles deeply and try to explain and the advantages and disadvantages and challenges which they encountered than other real time FEM simulators. In some of these articles, authors try not to just explain the sensors/phantoms/manipulators; however, they are trying to explore the different simulation opportunities. I would suggest authors to include these references and compare their simulation methods with their own. \
      1. Thank you for the considered references. We know the field is moving on to capture the combined effects of force and shape, but this was not included in this research. I plan to include some force and shape consideration findings in a generative design application late this year. In principle, force and shape consideration can be included in the reduced order model, albeit with increased simulations required to train the reduced order models. This is our plan, but for a later journal paper. I hope you understand. 
   3. In these following articles, authors are trying to develop some sensory systems using some affordable materials such as paper, double sided tape and Aluminum films, whereas they haven’t published some actuator designs. I believe, the major challenge is to explore different materials modelling in software while exploring the CAD based FEM modelling using COMSOL Multiphysics. I suggest authors to include these articles and explain the challenges faced by them and how could they encounter this. These tow above papers present the FEM of deformation in large deflection diaphragm whereas, the proper simulations are not performed.
      1. These five papers are quite interesting. Coming from an African university, I am always interested in topics that use local and inexpensive alternatives in cutting-edge research. I have made a note to contact the authors to see if they are interested in collaboration. I have read through all five papers, and as far as I can see, the common aspects are that they all look at applications with coupled electrical and mechanical components and use the COMSOL platform. I have looked into COMSOL and found that as a general-purpose structural FE modelling software, it has similar capabilities to other nonlinear FEM codes like LS-Dyna, Abacus and MSC.Marc so, I have not explored it in detail for this work. However, I see how the general field input used in the papers referred to could benefit from it. Especially considering that all the cases referred to are fully coupled multi-physics cases. The electromagnetic contribution to the structural analysis can explain some differences between the analytical and simulated models. I must admit that I don't fully see how computer-aided drawing is used or how this connects with the work presented in this paper. 
   4. This paper performs the simulation of capacitive change using CoventorWare tool whereas the simulations needed to scale down the cantilever size and thickness as well. How do authors explain that in this reputed journal?  
      1. Again, this refers to an interesting paper; however, I am not sure how it connects to the research we present here.
3. There are some research performed the finite element modelling of composite membrane based capacitive sensors; whereas, there are some soft actuators which design with composite layers to control the deformation and deflection. How could the authors claim that their modelling technique could be utilized to define the composite membrane, structure and deformation and other mechanical/electrical simulations.
   1. First I would like to address a point of clarity. The method presented here is better suited to the design of soft robots and actuators with shape changes defined as their behaviour, typified by multi-objective optimisation. The application of capacitive sensors, though complex, represents a single-step behaviour at the fundamental level of that research. The success of the ESP-based method proposed here would depend heavily on the choice of a reduce-order model representing the electro-mechanical behaviour of the sensors. Should a parametric model such as those used in the described papers replace the 3-unit FE model of this paper and the 2D mechanistic model representing the kinematics of the bending segment be replaced by a support vector machine representing the electromagnetic response, I think the method will still work. 

Thank you again for your comments. I know they have helped improve this paper. 

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper presents a mathematics method for conducting a generative design pipeline for soft robots to match with target deformation trajectory and grasping region. In particular, reduced-order models are used to predict the performance of the robot and an optimization-based solver is applied to get the design.

I have a few suggestions to help authors improve the manuscript:

1. In the introduction, please reformulate related work as a paragraph rather than using bully points.

2. The result presented in Fig.8 and Fig.11 looks good, I would like to also see the result to optimize for the locomotion performance of a 4-leg soft robot as mentioned in line 222 of page 4.

3. In the optimization pipeline, do you select the segmentation number as a fixed parameter or it can also be changed? If it’s the fixed case, how difficult it could be to optimize it together with the shape?

4. Please briefly discuss how the proposed method can work for 3D cases and if it will bring a significant increment in computational cost.

Author Response

Thank you for your review. Please see my comments below. I have organised your comments in an enumerated list below. With my comments italicised below each. 

1. In the introduction, please reformulate related work as a paragraph rather than using bully points.

   1. We have reformatted the introduction into a more traditional paragraph and made major changes to the style of referencing. We have now favoured fewer, more targeted references and the inclusion of a few more general reviews. Please see the attached marked-up version of the paper for comparison. 

2. The result presented in Fig.8 and Fig.11 looks good, I would like to also see the result to optimize for the locomotion performance of a 4-leg soft robot as mentioned in line 222 of page 4.

   1. It is indeed our intention to apply this method to the problem of locomotion. In the current framework, we focus solely on the deformation alone and are in the process of including the combination of shape change and external loading into an updated reduced order model and should have some results in a few months that are worth publishing. Our intention with this paper is to qualitatively compare the method to a previously established optimisation case and qualitatively evaluate extending the method to a more complex movement control problem. The inclusion of the external load is out of scope for this paper. 

3. In the optimization pipeline, do you select the segmentation number as a fixed parameter or it can also be changed? If it’s the fixed case, how difficult it could be to optimize it together with the shape?

   1. I am unsure which segment you refer to, so I will address two cases. 1) In the case of partial curve matching, we make use of analytical functions for the target curve, while the reduced order model makes use of discrete segments. We use the number of segments in the reduced order model to determine the number of segments in the PCM algorithm. 2) Regarding the number of wedge segments, we leave this open to the optimizer. When we decided to allow the optimiser to select from a list of multi-unit groups, we lost the ability to dictate the number of segments in the final design implicitly. We could, of course, enforce this with a constraint, but we never found that to be necessary for our examples.

4. Please briefly discuss how the proposed method can work for 3D cases and if it will bring a significant increment in computational cost.

   1. The method could be expanded into a 3D space without much change to the conceptual framework, but we would need to change the reduced-order model. Currently, the reduced order model is defined and trained for two degrees of freedom, and at least one more would need to be included to account for movement in an additional dimension. For example, an additional angle change can be defined for the reduced order model, and a series of FE simulations with movement in 3D can be generated for training. The optimisation routine and partial curve matching will remain the same. I do expect a meaningful increase in computational cost, but this will be predominantly limited to the training of a reduced-order model. Once this is trained, the optimisation should still run in a few seconds. I have added this example to the conclusion to further highlight the framework's scalability. Please see the attached marked-up version of the paper. 

Thank you for the insightful comments. I hope my feedback and marked-up report address your concerns. 

Author Response File: Author Response.pdf

Reviewer 3 Report

The author presented an optimization design method for soft robot . This method is expecting to be applied to pneumatic soft grippers. But the manuscript is need reorganized. The major issue is:

    The author uses the ESP framework to generate many examples of pneumatic soft robot structures, but does not prove that these generated structures have better performance.

    The reviewer believes that at least some simulation comparisons must be provided. To prove that this technical framework is superior to traditional design methods.

Author Response

Thank you for your comments please see my responses in the list below. 

1. The author uses the ESP framework to generate many examples of pneumatic soft robot structures, but does not prove that these generated structures have better performance.
   1. Thank you for highlighting this as unclear in the paper. We replicate a previous paper to establish the relative performance of the proposed reduced order model and ESP design framework. We then establish the method's utility in a designer-in-the-loop test case. I have adapted the introduction, abstract and conclusion to communicate this better. Please see the attached marked-up revision. 
2. The reviewer believes that at least some simulation comparisons must be provided. To prove that this technical framework is superior to traditional design methods.
   1. Thank you for highlighting that we have not communicated this well. The results section is broken into quantitative and qualitative sections. The first section quantitatively compares the proposed method to an established benchmark, while the second qualitatively evaluates the method on a new design task. I have adapted the introduction, abstract and conclusion to communicate this better. Please see the attached marked-up revision. 

Thank you again for your comments. 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

I congratulate authors to go through the manuscript once again and replying on the author’s comments very clearly. 

Apart from all the responses, I would still like to recommend that the authors need to go through the articles (which I have recommended in previous review report) which present the simulations using Comsol Multiphysics, CoventorWare, Sofa Framework and other simulators such as Ansys and other tools which authors also mentioned in the response letter. 

Otherwise, there are some review papers which authors need to cite before the final decision on the manuscript. There are some articles which authors have already cited and explained the introduction, material and result sections. Authors need to include the outlook section or can combine the Conclusion, outlook and Future work section in which they can explain the bending and deformation technique using FE modelling. They can refer the previously suggested references.

Ref:

1. G. Runge et al., A framework for the automated design and modelling of soft robotic systems. CIRP Annals, 66(1), 9-12, 2017.

2. R. B. Mishra, et al., "Recent progress on flexible capacitive pressure sensors: From design and materials to applications." Advanced materials technologies 6, no. 4 (2021): 2001023.

Find some other references as well and try to include them in outlook or discussion. Including that don’t forget to cite all the other included references in the journal format. Afterwards, as a reviewer, I will be able to recommend the article for further consideration.  

Author Response

Response to Reviewer 1 Comments

Thank you again for taking the time to review this manuscript and provide feedback. 

Point 1: Apart from all the responses, I would still like to recommend that the authors need to go through the articles (which I have recommended in previous review report) which present the simulations using Comsol Multiphysics, CoventorWare, Sofa Framework and other simulators such as Ansys and other tools which authors also mentioned in the response letter.

Response 1: 

Thank you for suggesting reviewing articles that present simulations using Comsol Multiphysics, CoventorWare, Sofa Framework, Ansys, and other tools mentioned in our response letter. I have read through the articles you recommended and revisited the product pages for Comsol, CoventorWare, and Sofa Framework. As someone who is familiar with these tools and has used several of them in other projects, I can say that they are all powerful and well-respected platforms.

However, based on my experience and evaluation, I do not see any clear benefits to using them instead of other general-purpose nonlinear simulation software tools commonly used in our field. While these tools have unique features and strengths, many other excellent simulators can provide similar capabilities.

While we appreciate your suggestion to discuss the limitations of simulation in soft robot design and current challenges in the field, our paper focuses on presenting an open framework suitable for any simulation type and software rather than providing a detailed discussion of the simulations themselves. Suppose we include the requested discussion of the limitations of simulating soft robotic components. In that case, it will distract from the manuscript's narrative, which is a framework for including simulation in the design process. 

Regarding the papers by Mishra and Hussain that you referenced, while we are aware of their work in the design of soft sensors, we do not believe that a detailed discussion of their research would add significant value to our manuscript, which focuses on presenting an open framework that can be used with any simulation type and software. Our manuscript is intended to provide a general tool for the community rather than being limited to a specific subfield. Nonetheless, we appreciate your feedback and will consider it as we revise our manuscript.

Perhaps if you could be more detailed on how you see a discussion of problem-specific simulations and their challenges fitting into this more general design framework, I would understand your intentions better. 

Point 2: Otherwise, there are some review papers which authors need to cite before the final decision on the manuscript. There are some articles which authors have already cited and explained the introduction, material and result sections. Authors need to include the outlook section or can combine the Conclusion, outlook and Future work section in which they can explain the bending and deformation technique using FE modelling. They can refer the previously suggested references.

1. G. Runge et al., A framework for the automated design and modelling of soft robotic systems. CIRP Annals, 66(1), 9-12, 2017.

2. R. B. Mishra, et al., "Recent progress on flexible capacitive pressure sensors: From design and materials to applications." Advanced materials technologies 6, no. 4 (2021): 2001023.

Response 2: 

I understand you want me to add a future work section to the paper. I have updated the paper to include a short future work paragraph highlighting that the next step in this work is to include more complex load-displacement behaviour. 

Regarding the two papers you suggest including in the manuscript. The work by Runge et al. shares some elements with the work presented in this manuscript. Their process of parametrized design, structural simulation, kinematic reduced order model, and metamodel is shared with our work. This manuscript, however, provides a framework within which several different approaches, including that presented by Runge et al. can be used in a single learning model, with the complex interaction of components and robot behaviour not well to the approach presented by Runge et al. I have included a reference to this paper in the introduction section. 

The paper by Mishra et al, though a little more recent, focuses specifically on details relating to the design of flexible pressure sensors. The paper is comprehensive but targeted at one specific application. Including this level of detail in a limited use case such as this does not make sense. Should we include this level of detail for soft sensors, we would need to do something similar for other applications, and the paper will shift from a research paper to a review.  

Point 3: Find some other references as well and try to include them in outlook or discussion. Including that don’t forget to cite all the other included references in the journal format. Afterwards, as a reviewer, I will be able to recommend the article for further consideration. 

Response 3: 

I am unsure what references you would like us to include in the discussion of our results. As it stands, the discussion focuses on the work done in the paper and does not lend itself to the inclusion of additional references. 

In addition to your suggestions, I have also made several small adjustments to the paper. 

Reviewer 3 Report

The author has revised the manuscript according the reviewer's comments.

Author Response

Response to Reviewer 3 Comments

Thank you again for taking the time to review this manuscript and provide feedback. 

Point 1: The author has revised the manuscript according the reviewer's comments.

Response 1: Thank you for the clear confirmation. I did use this review opportunity to adjust some sentences for clarity. 

Round 3

Reviewer 1 Report

1. As the authors show in Figure 8, the visual representation of bending actuators, this previously mentioned softwares are also capable of performing these representations of not. Because I have seen plenty of published manuscripts that perform these kinds of simulations and show similar representations. Therefore, the comparison of the FEM tools which use CAD design modeling and perform nonlinear simulations and comparison with this work would be interesting. Here I show one manuscript which performed some comparison. Reference: Xavier, M. S., Fleming, A. J., & Yong, Y. K. (2021). Finite element modeling of soft fluidic actuators: Overview and recent developments. Advanced Intelligent Systems, 3(2), 2000187.

2. In the second comment, I was looking at the authors' answers to check if there is the possibility of following similar design procedures if the bendable/flexible/stretchable devices or sensors are on the soft robots. In the recommended articles, the authors discuss the placement of sensors (sometimes thin or ultrathin) on robotic systems and perform the grasping. 

3. In the third comment - I wanted to check the possibility if there is the possibility of 2nd comment the authors should find some sensor designs and place them on the soft robot and then perform the evaluations. Otherwise, they should include their approach's constraints and explain future opportunities. 

Author Response

Thank you for taking the time to clarify the remaining points from the previous two reviews. We understand your comments better. Overall, the remaining three remarks are very broad and deserve a more detailed discussion than will fit into this paper. Nevertheless, we have tried our best to address each comment below.  

1. As the authors show in Figure 8, the visual representation of bending actuators, this previously mentioned softwares are also capable of performing these representations of not. Because I have seen plenty of published manuscripts that perform these kinds of simulations and show similar representations. Therefore, the comparison of the FEM tools which use CAD design modeling and perform nonlinear simulations and comparison with this work would be interesting. Here I show one manuscript which performed some comparison. Reference: Xavier, M. S., Fleming, A. J., & Yong, Y. K. (2021). Finite element modeling of soft fluidic actuators: Overview and recent developments. Advanced Intelligent Systems, 3(2), 2000187. 

   I agree. A rapidly growing collection of papers shows simulations of soft robots using FEA. I suspect we could extract data from published results for comparison or build reduced-order models and encapsulations. Still, we would need to invest an extensive amount of time in verifying this for publication. I want to make it clear that for this paper, we are not attempting to bring additional insights to the simulation or reduced-order modelling of bending actuators. These are active fields in their own right. We are only highlighting a framework that allows a designer to use a collection of existing tools in a structured way to develop new and insightful designs. We have included only sufficient detail in discussing these two components to show readers that the components have been correctly done. Although I believe a more detailed comparison of simulation and bending actuator modelling would be valuable, it is beyond the scope of this paper. 

   In my experience, simulating soft robots is difficult. For instance, the degree of material strain is well beyond anything I have seen in any other field, to the point where even the material models commonly used to represent elastomers start to break down (few models can reasonably exceed stretch ratios of 4 or 5). The inflated shapes, too, are susceptible to geometric deviations. While preparing our previous work (see Ellis, referenced in this manuscript), we noted that even a 0.05 mm deviation from the ideal geometry on the scale of our actuators leads to significant variations in the inflated shapes, which makes validating the FE Models with experimental results very challenging. With this in mind we have chosen to limit our comparison to this case, where we know the validation. 

   Further, there is significant value in an extension of the three paragraphs in the appendix by Xavier. Getting good results using any of the available simulation tools is challenging. If we, as a field, can establish best practices, it will significantly improve the quality of research done. That said, we have limited our discussion to the details of the FE modelling process as it is not central to the paper's narrative. In this case, we make use of FE simulation. Still, we could use another digital simulation or experimental data to build metarepresentations of the actuator behaviour. We have updated the paper conclusion to clarify that the selected tools are out of convenience and that many other tools and techniques can be used similarly.

   Further, in section 2.3, we have referenced Xavier again to highlight some of the challenges expected when using FEM to model soft robots, noting the content of the appendix to the paper. We have further clearly stated that even though we make use of MSC.Marc there are many suitable alternatives with similar capabilities.

2. In the second comment, I was looking at the authors' answers to check if there is the possibility of following similar design procedures if the bendable/flexible/stretchable devices or sensors are on the soft robots. In the recommended articles, the authors discuss the placement of sensors (sometimes thin or ultrathin) on robotic systems and perform the grasping.
   
   We believe this method can integrate sensors, sensor placement, and control elements relatively easily. The ESP framework is very flexible and allows for the simultaneous development of actuator geometry, sensor placement and control. To include these additional components, a designer needs to generate encapsulations for the sensors and adapt the virtual environment to include sensors and control in the optimisation loop. For example, changing the objective function from deviation from a target shape to a particular motion profile or behaviour, such as light avoidance. We have adjusted the "Conclusions" section to "Conclusions and Future Work" to account for additional discussion and extended the future work portion to include a specific note that sensors and control elements can be included, with references, and how the environment can be augmented to include behavioural objectives. 
   
   
3. In the third comment - I wanted to check the possibility if there is the possibility of 2nd comment the authors should find some sensor designs and place them on the soft robot and then perform the evaluations. Otherwise, they should include their approach's constraints and explain future opportunities.
   
   Per our response to point 2 above, sensors and sensor placement can be included with this method. However, additional work will need to be done to generate and validate the encapsulations of the sensors and control components. Although this would be interesting and valuable, we believe that it is best left for a future paper. 

Thank you again for your clarifications. I think the changes suggested have improved the clarity of the paper. As before, I have included a marked-up version of the paper highlighting the changes we have made.