Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors present a bio-inspired sensor-less feedback control framework for legged robots based on the Klann linkage mechanism. The approach utilizes an A-LSTM model to estimate motor angles from nonlinear current signals, which are then integrated with a PI controller to establish a closed-loop system for gait stability and autonomous recovery. However, the following issues should be considered to improve the readability and research accuracy of this manuscript.

 

1. To properly contextualize the study’s contribution, the authors must expand the literature review to cover existing sensor-less control paradigms in legged robotics more rigorously. The current introduction lacks a comparison with state-of-the-art control strategies, such asAdaptive robust joint force control for rapid motion of hydraulic limb leg unit; Design and control for WLR-3P: a hydraulic wheel-legged robot; High dynamic position control for a typical hydraulic quadruped robot leg based on virtual decomposition control;
2. The manuscript currently lacks a detailed kinematic analysis and parametric specification of the utilized Klann Linkage.
3. The description of the proposed A-LSTM model is currently limited to qualitative textual explanations. To enhance theoretical rigor, it is recommended that this section be supplemented with explicit mathematical formulations of the attention mechanism, alongside pseudocode detailing the specific algorithmic architecture.
4. To substantiate the authenticity and efficacy of the proposed method, visualization of the raw dataset and training convergence is essential. Please incorporate figures to demonstrate model convergence and the quality of the fit.
5. The experimental evaluation relies heavily on statistical summaries and lacks temporal resolution. It is critical to include time-series diagrams that illustrate system dynamics over time.
6. Please ensure that all graphical representations are clearly annotated with appropriate physical units and precise time scales on the axes to maintain scientific rigor and interpretability.
7. Given the manuscript’s explicit premise that the attention mechanism mitigates specific limitations of standard LSTM, a direct empirical comparison is requisite to substantiate this claim, rather than limiting the evaluation to open-loop or rule-based models.
8. While the conclusion summarizes the quantitative results, it falls short in providing a qualitative synthesis of the core principles.

Author Response

Comment1: To properly contextualize the study’s contribution, the authors must expand the literature review to cover existing sensor-less control paradigms in legged robotics more rigorously. The current introduction lacks a comparison with state-of-the-art control strategies, such as Adaptive robust joint force control for rapid motion of hydraulic limb leg unit; Design and control for WLR-3P: a hydraulic wheel-legged robot; High dynamic position control for a typical hydraulic quadruped robot leg based on virtual decomposition control;
Response1: Thank you for pointing this out. We agree with this comment. Therefore, we have expanded the literature review to encompass the existing sensor-less control paradigms more rigorously. Specifically, we have reviewed the state-of-the-art control strategies from the papers you kindly suggested and added an objective analysis comparing these methods with our proposed approach to clearly highlight our study's contribution. This change can be found in the revised manuscript on page 1  of 20, lines 23 to 38.

Comment2: The manuscript currently lacks a detailed kinematic analysis and parametric specification of the utilized Klann Linkage.
Response2: We appreciate the reviewer’s valuable feedback regarding the kinematic modeling and parametric specification. We would like to clarify that the detailed kinematic analysis and the optimization of the Klann linkage parameters were the core focus of our foundational preliminary study. This prior work, which has been recently accepted for publication (Choi, W.; Jung, H.; Woo, S.; Park, S.; Bae, W.G. AI-Based DC Motor Angle Prediction Process for Endpoint Position Control of a Klann Linkage. Journal of Electrical Engineering & Technology 2026), established the optimized linkage ratios and trajectory modeling required to determine the most suitable control framework for this specific robotic platform.
In response to the reviewer’s suggestion, we have now incorporated this newly accepted study as a key reference and included a summarized version of the kinematic analysis and parametric specifications in Section 2.1.1 (Kinematic Modeling) of the revised manuscript. This addition ensures methodological continuity and provides the necessary technical details for the reader to understand the mechanical constraints of the linkage system. We believe this clarification and the added reference fully address the reviewer's concern.

Comment3 : The description of the proposed A-LSTM model is currently limited to qualitative textual explanations. To enhance theoretical rigor, it is recommended that this section be supplemented with explicit mathematical formulations of the attention mechanism, alongside pseudocode detailing the specific algorithmic architecture.
Response3 :We sincerely appreciate the reviewer’s suggestion to enhance the theoretical foundation of the manuscript. We agree that a formal mathematical description and a detailed architectural overview are essential for the reproducibility and rigor of the proposed A-LSTM model.
To address this, we have added a new section, 2.2.1. LSTM Cell Dynamics and Attention Mechanism. This section provides:

1. Mathematical Formulations: We have explicitly defined the LSTM gate dynamics (input, forget, and output gates) and the attention layer's computational flow, including the calculation of alignment scores, attention weights, and context vectors.
2. Algorithmic Architecture: A new figure (Figure [3]) has been included to illustrate the detailed data flow and the integration between the LSTM layers and the attention mechanism.

We believe these additions provide the theoretical depth and clarify the specific algorithmic contributions of our work.

Comment4 : To substantiate the authenticity and efficacy of the proposed method, visualization of the raw dataset and training convergence is essential. Please incorporate figures to demonstrate model convergence and the quality of the fit.
Response4 : We sincerely appreciate the reviewer’s constructive suggestion to include data visualization and training convergence analysis. We fully agree that such visual evidence is essential for demonstrating the transparency and reliability of the model's performance. To address this, we have added two new sections into the revised manuscript.
First, Section 3.2.1 (Dataset Visualization and Correlation Analysis) was added to insert raw current data and motor angle visualization data used in the experiment. This made it possible to clearly check the characteristics and noise level of the data according to the change in the walking cycle of the robot. In addition, we tried to secure the justification for the selection of the ALSTM model by adding a comparison graph of the models for application to the robot.
Second, Section 4.2.2 (Evaluation of Prediction Quality and Fit) has been newly established to describe the learning and validation process of the model in detail. We added a graph that visually compares the actual measured current values with those predicted by the A-LSTM model. This quantitatively shows that the proposed model maintains high prediction accuracy and quality even for complex nonlinear data.
We believe these visual supplements provide a robust empirical foundation, demonstrating how effectively the proposed A-LSTM model learns and responds to the dynamic load conditions encountered in real-world environments.

Comment5 : The experimental evaluation relies heavily on statistical summaries and lacks temporal resolution. It is critical to include time-series diagrams that illustrate system dynamics over time.
Response5 : We sincerely appreciate the reviewer’s constructive suggestion regarding the inclusion of time-series diagrams. We agree that such data could provide a more granular view of the robot’s transient behavior.
However, the primary focus of this research was the implementation and validation of the AI model’s real-time operation on a resource-constrained MCU, specifically the Raspberry Pi Pico. To ensure the real-time deterministic behavior of the control loop and to prevent detrimental latency that could arise from wireless data transmission or high-frequency storage operations, we intentionally prioritized system stability during the experimental phase.
Consequently, our data collection was focused on macro-level indicators (statistical summaries), which we believe effectively represent the overall robustness and integrated performance of the proposed proprioceptive loop in diverse environments.
We acknowledge this as a technical limitation of the current hardware setup. In our future work, we plan to incorporate more advanced data-logging frameworks and high-performance hardware to capture high-resolution temporal dynamics without compromising control stability. We hope the current statistical analysis sufficiently demonstrates the effectiveness of our proposed methodology.

Comment6 :Please ensure that all graphical representations are clearly annotated with appropriate physical units and precise time scales on the axes to maintain scientific rigor and interpretability.
Response6 : We sincerely appreciate the reviewer’s meticulous feedback regarding the clarity and scientific rigor of our graphical representations. We agree that incorporating precise units and time scales is essential for the accurate interpretation of the results, and we have updated the manuscript accordingly.
First, we have added specific annotations and physical units to the distance and time axes in all existing box plot graphs to improve data readability. This ensures that the quantitative values of the experimental results can be clearly identified by the reader.
Second, to provide an intuitive estimation of the robot's physical dimensions, a scale bar has been newly inserted into the actual photographs of the robotic platform.
Third, we have explicitly specified the detailed length ratios of the Klann linkage in the manuscript and relevant tables. This information serves as a critical foundation for understanding the robot’s overall size and its kinematic locomotive characteristics.
We believe these enhancements significantly improve the transparency of the experimental data and increase the interpretability of the hardware system presented in this study.

Comment7 : Given the manuscript’s explicit premise that the attention mechanism mitigates specific limitations of standard LSTM, a direct empirical comparison is requisite to substantiate this claim, rather than limiting the evaluation to open-loop or rule-based models.
Response7 : We fully agree with the reviewer that justifying the choice of the Attention-LSTM (A-LSTM) over a standard LSTM is crucial for the manuscript's integrity.
To address this, we have incorporated our recently accepted prior study (****) as a foundational reference for the model selection process. In that study, we conducted a rigorous comparative analysis of various deep-learning architectures, including standard LSTM and CNN-LSTM, using Circular MSE as the loss function to account for the periodic nature of motor angles.
The results from our prior empirical tests confirmed that A-LSTM achieves the lowest error rate and superior predictive stability during load fluctuations compared to the other models. Based on this established superiority, the primary objective of the current manuscript is not to re-benchmark different neural architectures, but to focus on the real-time implementation and robustness of the pre-selected optimal model (A-LSTM) on a physical robotic platform.
We have clarified this rationale in the revised manuscript in Section 3.2.1 (Dataset Visualization and Correlation Analysis). By referencing the previous comparative validation, we ensure that the methodological rigor is maintained without redundant experimental overlap. We believe this clarification substantiates our claim regarding the selection of the attention mechanism.

comment 8 : While the conclusion summarizes the quantitative results, it falls short in providing a qualitative synthesis of the core principles.
response 8 : Thank you for pointing this out. We agree with this comment. Therefore, we have thoroughly revised the conclusion section to provide a deeper qualitative synthesis. Specifically, we have described the underlying phenomena demonstrated by each experimental metric. Furthermore, by synthesizing the results across all experiments, we have added a comprehensive discussion on the core principles and the broader significance of these numerical findings. This change can be found in the revised manuscript on page [19 of 20], lines [514 to 527].

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript proposes an AI-based sensorless feedback control framework for a Klann linkage walking robot. The authors use motor current signals as internal sensory information and train an Attention-LSTM model to predict motor angles. The predicted angles are then integrated into a PI controller to achieve closed-loop gait control without encoders or IMUs.

My comments are as belows;

1) These should be explained clearly; What is fundamentally new compared to prior sensorless motor control studies? How this approach differs from existing current-based state estimation methods?

2) The dataset is described as 120,000 points but critical information is missing like total recording duration , Number of gait cycles recorded, Train/validation/test split ratio

3) Whether test data were collected under different environmental conditions ?

4) The manuscript does not justify the system operation frequency. For dynamic locomotion control 10 Hz is relatively low.

5) All experiments appear to be on flat indoor tiles, with no disturbance injection beyond AI.  Thus claims about real-world adaptability should be justified. 

6)  Some claims are overstated, like "This presents a new control paradigm,,"

7) Figures 3-5 are shown but detailed variance statistics are not provided

 

Author Response

Comment1 : These should be explained clearly; What is fundamentally new compared to prior sensorless motor control studies? How this approach differs from existing current-based state estimation methods?
Response1 : We agree that the distinction from existing studies needed more clarity. Unlike conventional sensorless control that relies on the physical modeling of current-torque correlations or complex analytical frameworks (e.g., VDC or ARC), our approach introduces a data-driven, bio-inspired 'proprioceptive loop.' We have redefined the motor current not merely as a variable for torque estimation, but as a cognitive feedback signal for real-time postural awareness. By utilizing the Attention-LSTM (A-LSTM) model, we effectively handle the nonlinearities of the Klann linkage without the heavy computational burden of traditional model-based strategies. This fundamental shift toward bio-inspired self-perception has been further emphasized in the revised Introduction (pp. 1-2).

Comment2 : The dataset is described as 120,000 points but critical information is missing like total recording duration , Number of gait cycles recorded, Train/validation/test split ratio Response2 : We appreciate the reviewer’s constructive feedback regarding the detailed specifications of our dataset. We agree that providing these metrics is essential for ensuring the reproducibility and statistical clarity of our study. In response to this comment, we have supplemented the "Experiments and Results" section with the missing information. Specifically, we have clarified that:

1. Total Recording Duration: The 120,000 data points, collected at a sampling frequency of 10 Hz, correspond to a total recording duration of approximately 200 minutes.
2. Number of Gait Cycles: Given the robot’s average gait cycle of 3.2 seconds, the dataset encompasses approximately 3,750 complete gait cycles, capturing a diverse range of load variations.
3. Data Split Ratio: The dataset was partitioned into training, validation, and testing sets using an 8:1:1 ratio (96,000, 12,000, and 12,000 points, respectively) to ensure robust model evaluation.

These details have been added to Section 4 (Experiments and Results) on page 13, lines 364–374 in the revised manuscript.

Comment 3: Whether test data were collected under different environmental conditions? Response 3: Thank you for this valuable comment. We completely agree with the reviewer that testing under diverse environmental conditions is essential for ensuring practical deployment and real-world robustness. In this study, the experiments were primarily conducted in a controlled flat-ground environment to rigorously validate the fundamental performance and stability of our proposed AI-based sensorless feedback control framework. As we noted in the manuscript, our primary objective was to “overcome engineering limitations and lay the groundwork for a sensing paradigm adaptable to complex terrains.” By establishing a reliable baseline through the biological principles of proprioception, we have secured the necessary foundation for further expansion. However, we fully acknowledge that the current lack of environmental diversity is a limitation. Following your insightful suggestion, we have added a discussion in the revised manuscript regarding this limitation and our plans for future research. This will involve testing the controller on irregular and slippery terrains to further evaluate its adaptability and robustness. We are grateful for this feedback, as it has greatly helped in clarifying the scope and future direction of our work. This change can be found in the revised manuscript on page [1 of 20], lines [2 to 15] and page [19 of 20], lines [504 to 513]< !-- notionvc: 78e5a06e-753c-4e3b-aae5-0b7fe2596102 -->

Comment4 : The manuscript does not justify the system operation frequency. For dynamic locomotion control 10 Hz is relatively low.
Response4 : We appreciate the reviewer’s feedback. We acknowledge the reviewer's observation that 10 Hz is lower than the frequencies typically used for high-speed dynamic locomotion. However, the choice of 10 Hz was an intentional engineering decision based on the specific objectives and hardware constraints of this study.
First, the primary goal of this research is to achieve sensorless control using low-cost, resource-constrained hardware. Utilizing a Raspberry Pi Pico requires a critical balance between the computational overhead of real-time A-LSTM inference and the sampling frequency of I2C-based current sensors.
Second, to maintain real-time deterministic control over a wireless socket connection, 10 Hz was found to be the optimal frequency to ensure minimal packet jitter and prevent buffer overflows, both of which are critical for the stability of the control loop.
Finally, considering the robot's average gait cycle of 3.2 seconds, a 10 Hz sampling rate provides 32 data points per cycle. Our experimental results confirmed that this temporal resolution is sufficient to capture the essential load signatures and mechanical interactions required for effective adaptive control.

Comments 5: All experiments appear to be on flat indoor tiles, with no disturbance injection beyond AI. Thus claims about real-world adaptability should be justified. Response 5: Thank you for this insightful critique. We recognize that our initial claims regarding "real-world adaptability" could be perceived as overstated, given that the experiments were primarily conducted on flat indoor tiles. As the reviewer correctly noted, this study is intended as a foundational step to validate the mechanistic effectiveness of the proposed AI-based sensorless control. Our primary goal was to assess the system's potential for adaptability by proving that a perception-control loop can be formed using only internal signals. We now realize that asserting broad "adaptability" without extensive unstructured terrain testing could be problematic, and we have revised the manuscript to more accurately reflect the scope of this baseline validation. However, to justify the potential of our system, we have further clarified that we did indeed inject physical disturbances in the form of "stuck states." This was done to verify the robot’s recognition and recovery capabilities under severe physical constraints, even within a flat environment. We have now reframed our conclusions to emphasize this as a "critical engineering cornerstone" for future real-world applications rather than a completed proof of universal adaptability. This change can be found in the revised manuscript on page [12 of 20], lines [347 to 357] and [13 of 20], lines [377 to 396].< !-- notionvc: c77bff51-d88c-4122-b76e-80e2f00c9616 -->

Comment6 : Some claims are overstated, like "This presents a new control paradigm,,"
Response6 :Thank you for pointing out some expressions that overemphasized the value of this study. Including the 'new control paradigm' you pointed out, the tone of the entire paper has been modified to convey the research results more objectively. We avoided subjective adjectives or definite expressions and focused on describing the experimental effectiveness of the proposed method.

Comment7 : Figures 3-5 are shown but detailed variance statistics are not provided
Response7 : We have addressed this by providing detailed statistical data for all experimental results. In the revised manuscript, we added Standard Deviation (SD) values to Tables 5, 6, and 7 to provide a complete picture of the control consistency. Additionally, we updated the captions for Figures 8, 9, and 10 to clarify the box plot distributions and whiskers, ensuring that the variance of the ten trials is transparently reported.

< !-- notionvc: f9c69174-6a1e-4690-91ed-df6296ff3e81 -->

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper, Bio-Inspired Proprioception for Sensorless Control of a Klann Linkage Robot using Attention-LSTM, addresses a relevant topic within the scope of biomimetics by proposing a bio-inspired mechanical system and its associated control strategy. The subject is timely and could be valuable for advancing biomimetic actuation and robotic implementation.

However, the paper currently has several structural and scientific shortcomings that limit its impact and clarity. The mechanical design is insufficiently detailed, especially concerning the 3D modeling process, geometric parameters, material selection, and design constraints. The absence of comprehensive CAD views, exploded assemblies, and dimensional justification makes it difficult to evaluate the system's originality and reproducibility. Additionally, although simulation tools are employed, the modeling framework is not transparent enough: key parameters, boundary conditions, validation procedures, and assumptions are not well documented, which weakens the methodological rigor.

The experimental validation is limited and does not substantiate the claimed performance improvements. Quantitative comparisons with existing biomimetic or robotic solutions are either absent or superficial, and the statistical analysis of the experimental data is unclear. Several figures have low resolution, small labels, and insufficient explanatory captions, which reduces their readability and interpretability. While the literature review covers some foundational works, it would benefit from stronger engagement with recent high-impact studies to better position the contribution within the current state of the art.

Overall, the paper demonstrates technical effort and conceptual potential. However, substantial revisions are required in terms of methodological depth, clarity of design presentation, quality of figures, and experimental validation before publication.

Author Response

Comment1 : The mechanical design is insufficiently detailed, especially concerning the 3D modeling process, geometric parameters, material selection, and design constraints. The absence of comprehensive CAD views, exploded assemblies, and dimensional justification makes it difficult to evaluate the system's originality and reproducibility.
Response1 : We appreciate the reviewer’s feedback regarding the necessity of a detailed mechanical description. We agree that providing comprehensive geometric and dimensional data is vital for assessing the reproducibility of the robotic system. In the revised manuscript, we have significantly expanded Section 2.1.1 (Kinematic Modeling) to provide a more rigorous description of the mechanical design. As our robotic platform utilizes the Klann linkage architecture—which was meticulously analyzed and optimized in our foundational prior study—we have now explicitly detailed the kinematic framework and the geometric optimization process within this section. Furthermore, to provide the requested dimensional justification, we have updated our graphical representations by incorporating physical scale bars into the photographs of the robot. By providing these scales alongside the specific linkage length ratios now listed in the text, we have ensured that the physical dimensions and proportions of the system can be accurately estimated by the reader. These additions clarify the design constraints and substantiate the mechanical originality of the platform while ensuring high reproducibility for future research.

Comment2 : Additionally, although simulation tools are employed, the modeling framework is not transparent enough: key parameters, boundary conditions, validation procedures, and assumptions are not well documented, which weakens the methodological rigor.
Response2 : We completely agree with the reviewer that transparency in the modeling framework is critical to maintaining methodological rigor. To address this issue, we have compiled a new section, 3.4 modeling framework and assumptions, to provide a centralized and detailed documentation of the simulation environment.
In this section, we tried to significantly improve the readability and transparency of our methodology by describing the boundary conditions of the robot and organizing various experimental assumptions. We hope these additions will provide the theoretical depth and rigor requested by our reviewers.

Comment3 : The experimental validation is limited and does not substantiate he claimed performance improvements. Quantitative comparisons with existing biomimetic or robotic solutions are either absent or superficial, and the statistical analysis of the experimental data is unclear. 
Response3 : We sincerely appreciate the reviewer’s perspective on the need for broader comparative validation. We would like to clarify that the primary focus of this manuscript is the practical implementation and feasibility of an AI-driven proprioceptive loop under extreme hardware and cost constraints.
This study functions as a pilot project to explore how a high-level deep learning model (A-LSTM) can be compressed and deployed on a low-cost MCU (Raspberry Pi Pico) using only a single current sensor. Unlike conventional high-end robotic studies that utilize expensive IMUs or encoders, our goal was to determine the "lower bound" of hardware required for autonomous recovery. Consequently, direct quantitative comparisons with state-of-the-art (SOTA) systems using premium sensors are limited, as they operate in entirely different cost and complexity domains. To address concerns regarding statistical clarity and proof, we have strengthened our internal comparative analysis. Perform paired t-test in all experiments, and report the standard deviation (SD) for each test, hope the correction will meet the needs of the reviewer.
We believe that these strengthened statistical indicators, combined with the real-time deployment on a resource-constrained platform, provide a empirical foundation for our claims. We have updated Section 4.1 - 4.3 to reflect these statistical enhancements and to more clearly define the exploratory scope of this study.

Comment4 : Several figures have low resolution, small labels, and insufficient explanatory captions, which reduces their readability and interpretability. 
Response4 : We agree that high-quality visual representations are critical for the effective communication of research findings. We appreciate the reviewer’s feedback on this matter and have taken the following steps to enhance the visual clarity and interpretability of the manuscript:
Enhanced Captions: We have expanded the explanatory captions for Figures 8, 9, and 10 to provide more comprehensive context and detailed descriptions of the observed trends.
Improved Label Visibility: The font sizes for all axis labels, legends, and annotations have been increased to ensure legibility throughout the manuscript.
Inclusion of Physical Units: Consistent with our response to previous comments, we have ensured that graphical representations now include precise physical units (e.g., seconds, meters) for improved interpretability.
We believe these modifications greatly enhance the overall readability of the study and ensure that our experimental results are presented with the necessary clarity.

Comment5 : While the literature review covers some foundational works, it would benefit from stronger engagement with recent high-impact studies to better position the contribution within the current state of the art. 
Response5 : Thank you for this constructive suggestion. We agree that a stronger engagement with recent state-of-the-art studies would better highlight our study’s contributions. Accordingly, we have significantly expanded the literature review to include high-impact studies on hydraulic legged robotics, such as the Fast Direct Adaptive Robust Control (FDARC), the hierarchical distributed control of WLR-3P, and the Virtual Decomposition Control (VDC) framework.
By analyzing these advanced model-based paradigms, we have explicitly positioned our work by contrasting their heavy reliance on high-performance sensor arrays and computational resources with our proposed AI-based sensorless framework. This comparison allows for a clearer demonstration of our research's contribution toward system lightweighting and structural simplification.
This change can be found in the revised manuscript on page 1 of 20, lines 23 to 38.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The Authors have addressed all of my concerns with the original manuscript. The revised manuscript is ready for publication

Author Response

Comment 1: The Authors have addressed all of my concerns with the original manuscript. The revised manuscript is ready for publication

Response 1: We would like to express our sincere gratitude to Reviewer 1 for the positive evaluation and for finding our revisions satisfactory. We appreciate the time and effort invested in reviewing our manuscript, and we are pleased that our responses and improvements have addressed all of the reviewer's concerns.

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript proposes a data-driven, bio-inspired sensorless control framework for a Klann-linkage walking robot. The authors employ an Attention-based LSTM (A-LSTM) model to reinterpret motor current as a proprioceptive feedback signal, forming an internal perception-control loop without additional physical sensors. The experimental validation is conducted on a hardware platform operating at a 10 Hz sampling frequency.

In this revision, the authors have clarified dataset details, reduced overstatements, added statistical variance analysis, and refined claims regarding environmental adaptability.

The manuscript is suitable for publication in its present form, as the authors have adequately addressed all reviewer concerns.

Author Response

Comment 1: The manuscript is suitable for publication in its present form, as the authors have adequately addressed all reviewer concerns.

Response 1: We sincerely thank Reviewer 2 for the thorough review and for the positive final assessment of our manuscript. We are gratified that the clarifications regarding the dataset, the addition of statistical variance analysis, and the refinement of our claims on environmental adaptability were found to be sufficient. The reviewer’s constructive feedback during the initial round was instrumental in enhancing the technical rigor and clarity of this work. Thank you for recommending our paper for publication.

Reviewer 3 Report

Comments and Suggestions for Authors

The revised manuscript shows a significant improvement over the previous version. The authors expanded the kinematic and mechanical descriptions, clarified the modeling framework, strengthened the statistical analysis, improved figure readability, and broadened the literature review. Including additional details in Section 2.1.1 about the Klann linkage architecture and geometric ratios partially addresses concerns about mechanical transparency. Section 3.4, which is new, improves methodological clarity by addressing the modeling framework and assumptions. The addition of paired t-tests and standard deviation reporting enhances the credibility of the experimental evaluation. Figure captions and graphical readability have also improved.

The contribution of the manuscript is now clearer. It demonstrates the feasibility of deploying an AI-based proprioceptive recovery loop on extremely limited hardware with minimal sensing. However, the study is still exploratory and pilot in nature, rather than a fully matured research contribution with comprehensive validation. Also, there are relatively few bibliographical references.

Overall, the revision is a significant improvement over the original submission. However, the scientific depth, reproducibility documentation, and comparative validation could be strengthened.

Author Response

Comment 1: However, the study is still exploratory and pilot in nature, rather than a fully matured research contribution with comprehensive validation. 
Respond 1: We genuinely value Reviewer’s insightful perspective regarding the current stage and academic depth of our work. We agree that transitioning from an exploratory "pilot" phase to a more matured research contribution is a vital step for publication. 
To address the reviewer’s point regarding scientific depth, we have added a kinematic-based discussion in Section 4.3. We clarified how the A-LSTM attention mechanism correlates with the specific physical phases of the Klann linkage, providing a theoretical explanation for article. In addition, we tried to increase the persuasiveness of the characteristics of the paper felt by reviewers and readers by adding an explanation of the pilot nature of the paper to the conclusion.

Comment 2: Also, there are relatively few bibliographical references.
Respond 2: We sincerely thank the reviewer for the constructive feedback and for pointing out the need for stronger literature support. We completely agree that reinforcing the theoretical and biological foundations significantly improves the quality and persuasiveness of our manuscript.
Following your valuable suggestion, we have carefully reviewed our manuscript and incorporated seven highly relevant, state-of-the-art references. To preserve the original flow of the manuscript while maximizing the scientific rigor, these references were strategically added to explicitly support our existing claims regarding time-series AI modeling, biomimetic proprioception, and hardware design constraints.

Author Response File: Author Response.pdf