A Deep Learning-Based Framework for Offline Robotic Weld Path Generation Using a Single Top-View RGB-D Image
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors
This study presents a deep learning-based framework for weld seam extraction and 3D welding path generation using a single top view RGB-D image of a pipe workpiece. The overall approach adopted in the research holds practical engineering value. However, it exhibits numerous deficiencies in aspects such as the depth of innovation, the completeness of experiments, the logic of writing, and the comparison scores. Therefore, it necessitates substantial revision prior to re - examination.
- The logical structure of the introduction section is relatively weak. It is advisable to rewrite it. In the introduction, the author has repeatedly identified the flaws and deficiencies of the existing methods. Based on these identified deficiencies, the author should propose corresponding improvement measures in this research.
- The description of the innovation points lacks clarity. What is the core incremental value of the innovation?
- In the introduction, the four contribution points are described in terms of generalization, whereas the absence of quantitative comparison underscores incremental innovation.
- The "Materials and Methods" section is excessively long. It is advisable to streamline it. Since the deficiencies of the existing methods have already been mentioned in the introduction, there is no need to address them repeatedly in this section.
- Single-overview imaging exhibits inherent occlusion defects; however, this issue is only briefly mentioned in the discussion section of the paper. It is recommended that the author put forward some solutions to tackle this problem.
- It is recommended that the author perform a comparative analysis between this study and the prevailing methods of the same type.
- Only the offline path generation has been accomplished, and no online real - time weld seam tracking welding experiments have been carried out. The title of the paper indicates welding automation; however, there is a deficiency in verifying trajectory adaptive correction during the dynamic welding process, which is inconsistent with the industrial online welding requirements. It is recommended that the author make the requisite revisions.
- The RGB-D manual registration solely offers a set of fixed parameters. It is advisable that the authors elaborate on the parameter optimization process.
- Figure 5. The utilization of quartic polynomial fitting solely demonstrates the application of the least squares method, yet fails to analyze the rationale for selecting the order (specifically, why not the 3rd or 5th order). It is suggested that the author offer an explanation.
- Figure 9 only presents the average error of 8 sampling points. It fails to offer the error distribution of all test samples, the maximum error, and the standard deviation. Moreover, the data is not sufficiently reliable. An additional explanation is recommended.
- In the conclusion section, the primary findings of this study ought to be presented, and the discourse on future research directions should be located elsewhere.
Author Response
Thank you for taking the time to review this manuscript.
Please see the attachment containing the detailed responses.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for Authors
This paper explores the development of a compact and cost-effective approach to automated pipe welding based on weld detection and segmentation using a single image acquired by an RGB-D camera. The proposed method combines region-of-interest detection, segmentation, image post-processing, and weld trajectory generation, improving the accuracy of weld detection and generating a practical welding robot path. The relevance of this research stems from the growing industrial demand for reliable, affordable, and flexible robotic welding systems capable of handling complex product geometries without the need for expensive multi-sensor systems. The results confirm the potential of the proposed approach for improving the efficiency and accessibility of intelligent welding systems.
However, the paper requires further refinement.
1. The authors note that modern deep learning models require large training samples, but offer little analysis of whether their dataset of 1,476 images is sufficient for training deep learning models. Arguments regarding the representativeness of the sample and its statistical sufficiency should have been provided. (lines 57–79)
2. The data acquisition procedure is described in sufficient detail, but it remains unclear how exactly 100 frames were selected for depth averaging. No study of the effect of the number of frames on the final accuracy of trajectory construction is provided, so the chosen value appears empirical and insufficiently substantiated. (lines 135–147)
3. In describing the YOLOv8 model, the authors list the architectural features of the network, but provide virtually no justification for the choice of a specific model configuration. It would have been necessary to indicate which YOLOv8 variants (n, s, m, l, or x) were tested, how they differed in speed and accuracy, and why the particular variant used was ultimately selected. (lines 155–177)
4. The method for aligning RGB and depth images is based on manual tuning of several parameters. This approach can significantly depend on the specific operator and experimental setup. The reproducibility of the procedure and the ability to automatically determine image alignment parameters for different cameras and production conditions should be discussed. (lines 247–264)
5. The dataset created by the authors includes five types of pipes, but the range of geometric characteristics remains limited. Pipes with significantly smaller or larger diameters, different types of edge preparation, uneven gaps, and real-world manufacturing defects are not considered. This somewhat limits the applicability of the obtained results to a wide range of industrial problems. (lines 310–340)
6. The Results section logically continues the methodological section of the article and contains a large number of quantitative indicators. However, statistical processing of the experimental results is practically absent. All conclusions are based on single metric values ​​without confidence intervals, significance tests, or analysis of the scatter of results, which reduces the credibility of the experimental section of the article. (lines 341–376)
7. The results of the detection model demonstrate very high Recall and mAP50 values, but an analysis of erroneous cases is missing. It would be helpful to provide examples of false alarms and missed welds, and to explain the reasons for such errors under different pipe geometries and lighting conditions. (lines 343–355)
8. In analyzing the accuracy of trajectory construction, the authors demonstrate a significant reduction in errors after polynomial approximation. However, the impact of the resulting error directly on the quality of the welded joint is not verified. It would be significantly more convincing to present the results of actual welding followed by an assessment of the quality of the resulting weld. (lines 408–433)
9. The conclusion generally corresponds to the stated objective of the study and reflects the main results obtained. However, some conclusions are formulated much more broadly than supported by experiments. In particular, claims regarding the industrial applicability of the method require additional testing on a significantly larger number of product types and production conditions. (lines 489–505)
Author Response
Thank you for taking the time to review this manuscript.
Please see the attachment containing the detailed responses.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for Authors
This article is devoted to the development of a welding automation system. The paper describes the detection and segmentation of pipe welds using a single overhead image obtained with an RGB-D sensor. The paper describes a research methodology using a UR5e welding robot with an L515 camera. Some research results are presented.
Notes:
1. The article text does not indicate whether the pipe welding was performed using rotary welding or not.
2. There is no information on the welding modes, welding consumables, and equipment. This is important for welding.
3. There is no information on the edge preparation and weld joint geometry. This is important for V-groove welding. A deviation of 3.7 mm can lead to incomplete weld edge fusion.
4. There are no results from the analysis of the weld edge fusion (the macro and micro structure of the weld joint). This ensures the quality of the welded structure.
5. The conclusions do not contain information on the quality of the welded joint. Based on the presented research results, the stability of weld quality is not guaranteed.
The research described in the article addresses the pressing issue of welding process automation. However, the presented results are inconclusive and do not demonstrate the advantages of using the proposed method. The work requires significant revision.
Author Response
Thank you for taking the time to review this manuscript.
Please see the attachment containing the detailed responses.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for Authors
The author has carefully revised this manuscript in accordance with the prior review comments, and has also provided necessary explanations and clarifications for the relevant issues. The reviewer holds that the manuscript currently satisfies the acceptance criteria, and therefore recommends its acceptance for publication
Author Response
Thank you for your positive evaluation of our manuscript. We sincerely appreciate your valuable comments throughout the review process.
Although no further revisions were requested, we have attached a summary of the additional revisions made to the manuscript for your reference.
Please see the attached response document for details.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for Authors
Overall, the authors addressed the comments very well. For almost every point, they went beyond textual explanations, performing additional experiments, expanding sections of the article, adding new tables, figures, and discussions of the results. The revised version significantly strengthens the methodological section, experimental justification, and discussion of the method's limitations. Particularly positive is that the authors did not attempt to formally respond to the comments, but in many cases conducted additional research (comparison of YOLOv8 variants, study of the number of averaged frames, statistical processing of the results, and analysis of the method's applicability). Thanks to this, the article has become significantly more convincing and scientifically sound. In my opinion, approximately 85-90% of the comments were fully addressed, and the remaining issues are more likely to be improvements to the work than fundamental shortcomings. However, several comments remain incompletely addressed.
Comment 4. The authors described the manual calibration procedure and its reproducibility in detail, but they barely assessed the influence of operator bias. It would be useful to provide an experiment demonstrating the repeatability of the settings by different operators or by a single operator across multiple independent calibrations, as well as to quantify the spread of the obtained parameters and the resulting accuracy.
Comment 5. The authors correctly acknowledged the limitations of the dataset and moved the discussion to the limitations section, but the paper still lacks experimental confirmation of the method's ability to be generalized to new pipe types, edge preparations, and real-world manufacturing defects. A more detailed discussion of the expected limitations of the model and possible behavior when exceeding the training set would be desirable.
Comment 7. This comment is only partially resolved. The authors explained that they observed virtually no false positives or missed errors, but the lack of error examples does not fully address the request. Even with very high accuracy, it would be useful to show several of the most challenging cases with small bounding box offsets and briefly discuss the causes of these deviations.
Author Response
Thank you for taking the time to review this manuscript.
Please see the attachment containing the detailed responses.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for Authors
The article "A Deep Learning-Based Framework for Offline Robotic Weld Path Generation Using a Single Top-View RGB-D Image" is devoted to the development of a welding automation system. The paper describes the detection and segmentation of pipe welds using a single top-view image acquired with an RGB-D sensor. The paper describes a research methodology using a UR5e welding robot with an L515 camera. Some research results are presented.
Notes:
1. Line 358: Weld geometry is determined not only by welding current but also by arc voltage and welding speed. All welding mode parameters must be specified.
2. Page 18, Figure 12: The weld quality raises many questions. A photograph of the weld geometry in cross-section (macrosection) would be helpful.
The research described in the article addresses the current topic of welding automation. The article may be accepted for publication subject to correction of the comments.
Author Response
Thank you for taking the time to review this manuscript.
Please see the attachment containing the detailed responses.
Author Response File: Author Response.pdf
