Fabric Flattening with Dual-Arm Manipulator via Hybrid Imitation and Reinforcement Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper introduces a dual-arm fabric-flattening method based on a cascaded Proposal–Action network with a hybrid training framework. Experimental results demonstrate that the hybrid training framework substantially improves the overall flattening success rate compared with a policy trained only on human demonstrations. And we have the following comments:

1 The paper provides a structural diagram of AN. We could also create a structural diagram for PN to enhance the comprehensiveness of the method description and improve its readability.

2 The paper mentions torque sensors, but it does not provide a specific description of their functions. Could you please supplement some information on the functions of torque sensors and, at the same time, provide relevant data on the operation process of other robotic arm manipulations to facilitate readers' analysis and understanding?

3 The method employed for controlling the robotic arm during the process of flattening fabric pieces is mentioned in the paper but not elaborated upon. It is recommended to use hybrid force/position control as an approach.

4 The paper mentions that dual-arm systems are superior to single-arm systems. Could you please elaborate on this point and, ideally, supplement it with detailed comparative experimental data?

5 Researchers need to consider the issue of mutual interference during motion when the dual arms are operating. Could you please write down your solution to this problem?

   

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript introduces a dual-arm fabric flattening method based on a cascaded Proposal–Action network and a hybrid training framework that combines imitation learning from human demonstrations with reinforcement learning using real-world feedback. The approach is evaluated on baseline and variant fabrics of different thickness and stiffness, achieving higher one-shot and overall success rates than prior single-arm methods, while reducing the required number of flattening actions. The topic is relevant to deformable object manipulation and automated garment manufacturing. Overall, the manuscript is still missing improvements in baseline comparisons, evaluation metrics, methodological details, generalization analysis, and reproducibility as described below:

1. The key comparison is with the new method vs single-arm baselines. They are not comparing with the other two arm methods. Authors should add the recent two-arm methods under the same test rule. Authors should report one-shot rate, overall rate, and average action count with a 95% confidence range.
2. The success test, on the other hand, is sensitive to the fabric area from image masks. The authors do not explain what these mask steps are and what the cutoffs are. Authors should describe each step and what each cutoff represents. Authors should run the experiment with ±1% pixel error and see how the score changes. Authors should plot coverage versus action count with the average and the spread. Authors should also show some typical failures and the short reasons.
3. In the method section, the authors omit many important details. Projection Loss uses values that do not correspond to line segment length. Authors should describe what scale is needed or what division by length is needed. Authors should list what set values are for all loss weights. Authors should run ablations where one removes the Projection Loss, run others where there is only MSE. Authors should run ablation where they keep the PN decoder fixed instead of updating it.
4. The reinforcement learning training process lacks transparency regarding stability and reproducibility. Although exploration and exploitation datasets are reported, noise schedules, critic pretraining convergence, and overestimation mitigation strategies are not specified. The study should include training curves, stabilization techniques, and complete experimental settings including seeds, hardware, and impedance thresholds, to ensure reproducibility.
5. Generalization experiments on other fabric types are limited. It is not reliable to report only qualitative results without confidence intervals. It would be better to report one-shot success rate and overall success rate with confidence intervals, and further perform leave-one-type-out validation on zero-shot generalization. More challenging initial conditions should also be included. It is also recommended to discuss the limitations of bilateral symmetry assumptions.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This paper presents a novel approach for fabric flattening using a dual-arm robotic system, integrating imitation learning and reinforcement learning techniques. The main contributions include the development of a Proposal-Action Network that enhance the efficiency and effectiveness of fabric manipulation tasks, with experimental results demonstrating improved success rates through joint reinforcement learning. The strengths of the paper lie in its innovative methodology, comprehensive training regimen, and thorough evaluation across different fabric types.

The manuscript is generally clear, relevant to the scope of MDPI, and well-structured. The integration of reinforcement learning with robotic manipulation is timely and addresses significant challenges in automated garment manufacturing. However, there are areas that require further clarification and detail to enhance the overall scientific rigor of the study.

First, the authors should compare the proposed model with other baseline models of dual-arm fabric manipulation, as the current comparisons are only made against a single-arm benchmark model. While the manuscript presents promising results, a comparative analysis with dual-arm systems is essential for providing a comprehensive evaluation of its performance. By benchmarking against established dual-arm methods, the authors can effectively highlight the specific advantages of their model, including improvements in success rates, efficiency, and adaptability to various fabric types.

Second, the outstanding performance metric used in this paper is the "success rate," which is defined by the criterion that the relative coverage area of fabric exceeds 95% after flattening. However, it would be beneficial for the authors to provide further clarification on how this success rate is measured and validated. Specifically, details regarding the methodology for calculating the relative coverage area would enhance the understanding of the results. Additionally, it would be helpful to know how the 95% threshold was determined experimentally, including any trials or criteria that informed this decision. 

Third, the results section would benefit from presenting and evaluating the reward function values and Q-values associated with the reinforcement learning process.. Additionally, by discussing the relationship between these values and the observed success rates, the authors could better illustrate the mechanisms behind the enhancements achieved through joint reinforcement learning, thereby reinforcing the impact of their work on fabric manipulation tasks.

Additional comments are included in the annotations in the attached pdf version of this manuscript.

In the reviewer's opinion, this paper requires a major revision before being further considered for publication.

 

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

Please make sure that all Figures and Tables are placed appropriately within the text and referenced before their display. For example Figure 1 is present in section 1, but only referenced in section 3.

Figure 2 could benefit from an isometric view, rather than a flat projection.

Text in Figure 3 is too small and coordinate notations overlap with figure lines. Please increase the font and use a different color for coordinates and their text.

Throughout the text PN abbreviation is consistently used, but AN abbreviation is inconsistently exchanged with its full name.

Please update the citation Journal name in the GitHub repository. Also the file 'requirements.txt' is missing.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript is well-structured and features great scientific soundness, and can be considered for publication at MDPI Machines.