Human Action Generation from Skeleton Sequences: A Comparative Study of Mathematical and Bio-Inspired Algorithms

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a hybrid framework that combines mathematical modeling and bio-inspired optimization algorithms for human action generation from skeleton sequences. The following issues need to be considered.

1. Some descriptions in the paper appear to contain errors. For example, the statement "Therefore, MicroAG algorithm can be used in work [18]" is problematic, since Ref. [18] does not introduce the MicroAG algorithm at all.
2. It is suggested to add comparison results with state-of-the-art baseline methods. Meanwhile, the paper lacks quantitative analysis of the experimental results, which leads to insufficient support for the superiority of the proposed method.
3. The paper does not specifically explain how the homotopy curve parameters and the number of key poses affect joint motion. It lacks case analysis of the internal decision-making mechanism of the model, resulting in poor interpretability.
4. Is the proposed algorithm still applicable under noise interference and data missing conditions? Please supplement relevant explanations and verification. 

Author Response

This paper proposes a hybrid framework that combines mathematical modeling and bio-inspired optimization algorithms for human action generation from skeleton sequences. The following issues need to be considered.

1. Some descriptions in the paper appear to contain errors. For example, the statement "Therefore, MicroAG algorithm can be used in work [18]" is problematic, since Ref. [18] does not introduce the MicroAG algorithm at all.

 

Response:

 

We thank the reviewer for pointing out this ambiguity. The original statement was not intended to suggest that Ref. [18] introduces or uses the MicroGA algorithm. Instead, the intention was to indicate that, based on the results obtained in this work, the MicroGA algorithm could potentially be applied to problems similar to those addressed in Ref. [18]. The sentence in lines 426–427 has been revised and replaced with “The results obtained in this study suggest that the MicroGA algorithm may be suitable for application in problems similar to those addressed in [18].” to clarify this point and avoid misinterpretation. Additionally, the manuscript has been carefully reviewed to ensure consistency between the cited references and the statements made about them.

 

 

2. It is suggested to add comparison results with state-of-the-art baseline methods. Meanwhile, the paper lacks quantitative analysis of the experimental results, which leads to insufficient support for the superiority of the proposed method.

 

Response:

We thank the reviewer for this valuable suggestion. We agree that comparisons with state-of-the-art methods would provide additional context for evaluating the proposed approach. However, the primary objective of this work is not to demonstrate superiority over existing methods, but to assess the feasibility of generating motion transitions from skeleton sequences using a homotopy-based framework.

 

To provide a quantitative evaluation, the proposed method is analyzed using Dynamic Time Warping (DTW) as a measure of similarity between generated and reference sequences (Table 5). In addition, multiple runs of the optimization algorithms have been performed, and statistical descriptors, including mean, standard deviation, and 95% confidence intervals, have been incorporated to account for the stochastic nature of the methods, the information was was added in lines 306 to 409.

 

We acknowledge that direct comparisons with state-of-the-art approaches are not included in the current study. This limitation has been explicitly stated in lines 136-137 in the revised manuscript, and future work will focus on benchmarking the proposed framework against existing methods and incorporating additional evaluation metrics.

 

Overall, the revised manuscript provides a clearer and more balanced presentation of the contributions of this work as a feasibility study in lines 127-137 rather than a comparative performance analysis.

 

3. The paper does not specifically explain how the homotopy curve parameters and the number of key poses affect joint motion. It lacks case analysis of the internal decision-making mechanism of the model, resulting in poor interpretability.

 

Response

We thank the reviewer for this valuable observation.

 

We acknowledge that the original manuscript did not sufficiently explain how the homotopy curve parameters and the number of key poses influence the generated motion.

 

In the revised manuscript, we have included a qualitative analysis describing the effect of these parameters on the generated transitions. In particular, the number of key poses controls the level of detail in the trajectory representation: a higher number of poses allows capturing more complex motion patterns, while also increasing the dimensionality of the optimization problem. Conversely, a lower number of poses results in coarser transitions with more abrupt changes between configurations.

 

Furthermore, a new table has been included (Table 4) reporting the fitness of the best individual, Find, for each action and method, highlighting the role of λ as an indicator of transition completion.

 

These additions improve the interpretability of the model and clarify the relationship between the parameters and the resulting motion.

 

4. Is the proposed algorithm still applicable under noise interference and data missing conditions? Please supplement relevant explanations and verification. 

Response

We thank the reviewer for this important question. In the current study, the proposed framework is evaluated using skeleton sequences without missing joints or significant noise. Therefore, the method assumes complete and consistent input data. This point is clarified in lines 90–92.

The homotopy-based approach generates transitions by continuously transforming the full set of joint positions between poses. As such, the presence of missing or highly noisy joint data may affect the stability of the generated transitions, since the method does not explicitly incorporate mechanisms for handling incomplete or corrupted inputs.

In scenarios involving noise or missing joints, a preprocessing stage would be required to filter the data or reconstruct the missing joint positions before applying the proposed framework. Techniques such as interpolation, smoothing, or model-based joint estimation could be integrated for this purpose.

We acknowledge that robustness to noise and incomplete data has not been explicitly evaluated in this work. This limitation has been clarified in the revised manuscript in lines 90–92.

A general revision of the document was carried out, and the discussion and conclusion sections were updated.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors present a hybrid framework combining homotopy-based animation with bio-inspired optimization techniaues. While the study is well-motivated and addresses an interesting problem in animation synthesis, the paper has several significant problems:

- The optimization objective is not rigorously justified. Fitness is defined based on the homotopy parameter lambda approaching 1, but there is no theoretical justification why this correlates with perceptual realism and it is unclear what the relationship is between lambda and motion quality.

- The use of dynamic time warping as a similarity metric is reasonable, but it is not integrated into the optimization objective. Also, there is no discussion of its limitations (sensitivity to spatial realism or joint constraints). Basically, there is a disconnect between optimization objective lambda and the evaluation metric DTW.

- The experimental design is one of the most significant weaknesses: only one sequence per action is used from the KARD dataset and total evaluation is effectively based on 18 samples. There is no variability across subjects, styles or noise, no train/test split or generalization analysis. The results may very well be the outcome of overfitting to individual sequences. 

- The evaluation methodology is not rigorous enough. DTW is the only quantitative similarity metric. While confusion matrices are used for classification of generated actions, the classification method is not clearly described, and it is unclear whether this indicates motion quality or simply separability. A perceptual user studey would be useful here, as well as an evaluation of biomechanical constraints and/or joint angle or physical plausability analysis. As it stands, the claim of "realistic" motion is weakly supported.  

- Although multiple runs are mentioned, there is no statictical significance testing. No confidence intervals or hypothesis testing are reported. Standard deviation is shown, but not properly interpreted. Considering the stochastic nature of the algorithms, this is a major omission.

- The encoding scheme raises some concerns: only 4 bits for key poses, which means a maximum of 15 poses. Radius and increment are discretized with fixed scaling factors. The arbitrary discretization may limit the quality of the solution, and there is no justification for parameter ranges. 

- The authors claim that "the framework offers an effective solution for reconstructing realistic human actions", but the evaluation is limited and narrow, there is no comparison with state-of-the-art methods and there is no perceptual validation. Thus, the authors' claims should be signifcanly toned down. 

While I commend the authors for their effort, I consider that the improvements required for this to be a proper paper exceed even the time and scope of a major revision. I therefore recommend that the authors submit an entirely new paper where the above issues have been properly addressed.

 

Author Response

The authors present a hybrid framework combining homotopy-based animation with bio-inspired optimization techniaues. While the study is well-motivated and addresses an interesting problem in animation synthesis, the paper has several significant problems:

1- The optimization objective is not rigorously justified. Fitness is defined based on the homotopy parameter lambda approaching 1, but there is no theoretical justification why this correlates with perceptual realism and it is unclear what the relationship is between lambda and motion quality.

Response

We thank the reviewer for this important observation. In lines 84–92, the manuscript was revised to clarify the role of the homotopy parameter λ\lambdaλ and its relationship with the optimization objective.

In the proposed framework, \lambda is not intended to measure perceptual realism or motion quality. Instead, it is defined within the homotopy formulation as a parameter that represents the progression of the transition from the initial pose (\lambda = 0$) to the final pose (\lambda = 1. Therefore, \lambda serves as an indicator of transition completion. Values of \lambda close to 1 indicate that the generated trajectory successfully reaches the target pose, while lower values correspond to incomplete transitions.

Based on this interpretation, the fitness function is designed to favor solutions that achieve $\lambda$ values close to 1, ensuring that valid and complete transitions between poses are obtained. This formulation provides a clear and well-defined optimization objective based on transition feasibility and completion, rather than perceptual criteria.

To further address this concern, the manuscript has been revised to explicitly separate the optimization objective from the evaluation stage. While the optimization process focuses on achieving transition completion, the similarity between generated and reference motion sequences is evaluated independently using Dynamic Time Warping (DTW) in a post-processing stage. This distinction avoids associating \lambda with perceptual evaluation and provides a clearer interpretation of both components.

A new subsection (Section 2.4) in lines 266- 285 has been added to formally define the fitness function and clarify the role of \lambda, along with corresponding revisions throughout the manuscript.

2- The use of dynamic time warping as a similarity metric is reasonable, but it is not integrated into the optimization objective. Also, there is no discussion of its limitations (sensitivity to spatial realism or joint constraints). Basically, there is a disconnect between optimization objective lambda and the evaluation metric DTW.

Response

We thank the reviewer for this valuable observation. The separation between the optimization objective and the evaluation metric is intentional in the proposed framework. The optimization process is designed to ensure transition feasibility and completion through the homotopy parameter \lambda, which measures whether a valid transition between the initial and final poses is achieved. In contrast, Dynamic Time Warping (DTW) is employed as an external similarity metric in a post-processing stage to compare the generated motion sequence with the reference data. This separation allows the optimization process to remain computationally efficient and focused on generating valid transitions, while sequence similarity is assessed independently. Integrating DTW directly into the optimization objective would significantly increase computational cost and could bias the search process toward sequence matching rather than ensuring feasible transitions between poses. To address the reviewer’s concern, the manuscript has been revised to explicitly clarify this design choice and to discuss the limitations of DTW. In particular, DTW measures temporal alignment and similarity between sequences, but it does not explicitly account for biomechanical constraints, joint limits, or perceptual realism. Therefore, DTW is used as a complementary evaluation metric rather than as part of the optimization objective. In lines 383-409 the use of DTW in this work and its limitations are clarified.

A new subsection (Section 2.4) in lines 266- 285 has been added to clarify the separation between optimization and evaluation, and to include a discussion of the limitations of DTW.

3- The experimental design is one of the most significant weaknesses: only one sequence per action is used from the KARD dataset and total evaluation is effectively based on 18 samples. There is no variability across subjects, styles or noise, no train/test split or generalization analysis. The results may very well be the outcome of overfitting to individual sequences. 

Response

We thank the reviewer for this important observation. We agree that the current experimental design does not aim to evaluate generalization across subjects, styles, or noise conditions.

The primary objective of this work is to assess the feasibility of the proposed framework for generating motion transitions from skeleton sequences using a homotopy-based formulation. For this purpose, a single representative sequence per action was selected from the KARD dataset, resulting in 18 distinct action types. This design allows us to evaluate whether the framework can generate valid transitions across a variety of motion categories, rather than to assess statistical generalization performance.

The optimization process is performed independently for each input sequence, and therefore the method does not involve a training phase in the conventional sense. As such, the notion of overfitting is not directly applicable, since the framework does not learn a model from a dataset but instead solves an optimization problem for each given sequence. To clarify this point, the manuscript has been revised to explicitly state the scope and limitations of the experimental design. We also acknowledge that evaluating robustness to variability, noise, and multiple subjects constitutes an important direction for future work.

In lines 342-346 “The KARD dataset contains multiple sequences for each action with similar initial and final poses. Since the objective of this work is to evaluate the feasibility of generating transitions between representative poses, a single sequence was selected for each action. This allows different types of motions to be analyzed while avoiding redundancy in similar pose configurations.”

4- The evaluation methodology is not rigorous enough. DTW is the only quantitative similarity metric. While confusion matrices are used for classification of generated actions, the classification method is not clearly described, and it is unclear whether this indicates motion quality or simply separability. A perceptual user studey would be useful here, as well as an evaluation of biomechanical constraints and/or joint angle or physical plausability analysis. As it stands, the claim of "realistic" motion is weakly supported.  

Response

We thank the reviewer for this observation. We clarify that the confusion matrices were not intended to represent a classification task. Instead, they were constructed based on a similarity analysis using Dynamic Time Warping (DTW). For each bio-inspired algorithm, 18 motion sequences were generated, one per action. Subsequently, the DTW distance between each generated sequence and all reference sequences was computed. Each generated sequence was then associated with the reference sequence that yielded the minimum DTW distance. The confusion matrices were constructed by assigning a value of 1 to the closest match for each case. However, we acknowledge that presenting these results as confusion matrices may lead to misinterpretation, suggesting a classification framework. Therefore, to avoid ambiguity, the confusion matrices have been removed from the revised manuscript. The evaluation has been refocused on DTW-based similarity analysis.

With the mathematical method and the bio-inspired algorithms, the objective is to generate complete transitions between pairs of key poses. For this purpose, the objective function is based on the λ value obtained for each pair of key poses. Therefore, lines 372–380 explain that Table 4 was added to show the fitness obtained for each action using the best individual. Subsequently, lines 383–395 present the similarity calculation using DTW. It can be observed that when the fitness of the individual is close to 1, the DTW distance is lower, indicating greater similarity between the generated action and the original one.

In lines 474–477, it is explained that future work will incorporate biomechanical constraints, evaluate multiple sequences per action, dynamically adapt the number of intermediate frames, and integrate the proposed homotopy-based formulation with deep learning methods to improve the realism of more complex actions.

5- Although multiple runs are mentioned, there is no statictical significance testing. No confidence intervals or hypothesis testing are reported. Standard deviation is shown, but not properly interpreted. Considering the stochastic nature of the algorithms, this is a major omission.

Response

We thank the reviewer for this valuable observation. To address this point, the statistical analysis of the results has been extended. In lines 396-409, in addition to reporting the mean and standard deviation of the DTW values, a 95\% confidence interval has been incorporated to better characterize the variability of the results, considering the stochastic nature of the optimization algorithms.

In Figure 11, the impact of the standard deviation is explained in lines 363-369.

While formal hypothesis testing was not included, the added statistical descriptors provide a clearer and more robust comparison between methods within the scope of this work.

6- The encoding scheme raises some concerns: only 4 bits for key poses, which means a maximum of 15 poses. Radius and increment are discretized with fixed scaling factors. The arbitrary discretization may limit the quality of the solution, and there is no justification for parameter ranges. 

Response

We thank the reviewer for this insightful observation. The encoding scheme was designed to provide a compact and computationally efficient representation of the optimization variables, while maintaining sufficient flexibility to generate valid motion transitions.

The use of 4 bits for encoding the number of key poses allows representing up to 15 poses, which was found to be sufficient for capturing the main structure of the evaluated actions. Increasing the number of key poses would expand the search space and computational cost, without necessarily improving the feasibility of the generated transitions within the homotopy-based framework.

Similarly, the discretization of the homotopy parameters (radius and increment) was introduced to bound the search space and facilitate the optimization process. The selected ranges were defined based on preliminary empirical observations to ensure stable convergence of the homotopy tracing procedure.

We acknowledge that the discretization may limit the exploration of the parameter space. However, the objective of this work is to evaluate the feasibility of the framework under a controlled and tractable optimization setting. Extending the approach to continuous parameter representations or adaptive encoding schemes constitutes an interesting direction for future work.

Lines 249–255 were added to clarify this point.

7- The authors claim that "the framework offers an effective solution for reconstructing realistic human actions", but the evaluation is limited and narrow, there is no comparison with state-of-the-art methods and there is no perceptual validation. Thus, the authors' claims should be signifcanly toned down. 

Response

We thank the reviewer for this important observation. We agree that the original wording may overstate the scope of the proposed framework.

The manuscript has been revised to moderate the claims and better reflect the actual contributions of the work. In particular, references to "reconstructing realistic human actions" have been replaced with more precise descriptions focused on generating feasible and complete motion transitions between key poses.

Furthermore, the scope of the evaluation has been clarified to emphasize that the objective of this work is to assess the feasibility of the proposed framework, rather than to perform a comprehensive comparison with state-of-the-art methods or to evaluate perceptual realism.

A general revision of the document was carried out, and the discussion and conclusion sections were updated.

Reviewer 3 Report

Comments and Suggestions for Authors

This paper presents a hybrid framework combining homotopy-based mathematical modeling with four bio-inspired algorithms (GA, MicroGA, PSO, ACO) to generate human actions from skeleton sequences. It optimizes key pose selection and homotopy curve parameters. Evaluated on 18 actions from the KARD dataset, MicroGA and PSO produced the most realistic animations, with MicroGA preferred for its efficiency and consistency.

1. How is the threshold of λ > 1.1 for fitness penalty justified? Please clarify the relationship between λ values exceeding 1.0 and the quality of generated transitions. Refer to the paper: Intelligent disassembly scenario understanding for human behavior and intention recognition towards self-perception human-robot collaboration system
2. The confusion matrices for PSO and MicroGA appear identical. Please verify whether this reflects actual performance or a potential data entry error. Refer to the paper: Dynamic task allocations with Q-learning based particle swarm optimization for human-robot collaboration disassembly of electric vehicle battery recycling
3. Why was the number of intermediate frames fixed at 30? Could varying this number per action or pose pair improve motion naturalness?
4. The method selects key poses via k-means. How sensitive are results to k-means initialization? Were multiple runs performed to ensure stability?
5. Please clarify how the homotopy approach handles occluded or missing joints in skeleton data, and whether joint correlations are preserved during transitions. Refer to the paper: Multi-scenario digital twin-driven human-robot collaboration multi-task disassembly process planning based on dynamic time petri-net and heterogeneous multi-agent double deep Q-learning network

Author Response

This paper presents a hybrid framework combining homotopy-based mathematical modeling with four bio-inspired algorithms (GA, MicroGA, PSO, ACO) to generate human actions from skeleton sequences. It optimizes key pose selection and homotopy curve parameters. Evaluated on 18 actions from the KARD dataset, MicroGA and PSO produced the most realistic animations, with MicroGA preferred for its efficiency and consistency.

1. How is the threshold of λ > 1.1 for fitness penalty justified? Please clarify the relationship between λ values exceeding 1.0 and the quality of generated transitions. Refer to the paper: Intelligent disassembly scenario understanding for human behavior and intention recognition towards self-perception human-robot collaboration system

Response

We thank the reviewer for this important question. Section 2.4 was added in lines 266–285, where it is explained why λ should remain within the interval [0,1]. In lines 298–303, it is explained that if λ is greater than 1, the generated trajectory exceeds the target pose, and the manuscript specifies how an individual is penalized when the value of λ is greater than 1.1.

 

In the homotopy-based formulation, the parameter $\lambda \in [0,1]$ represents the progression of the transition from the initial pose ($\lambda = 0$) to the final pose ($\lambda = 1$). Therefore, $\lambda = 1$ indicates that the generated trajectory has successfully reached the target pose.

 

Values of $\lambda$ greater than 1 correspond to extrapolation beyond the final pose, meaning that the transition overshoots the target configuration. This behavior is typically associated with numerical instability or divergence in the homotopy tracing process, rather than a valid or meaningful motion transition.

 

For this reason, a penalty is introduced when $\lambda$ exceeds a threshold slightly above 1 (set to 1.1 in this work), allowing a small tolerance for numerical fluctuations while discouraging solutions that significantly overshoot the target pose.

 

This design ensures that the optimization process favors solutions that reach the final pose without divergence, promoting stable and complete transitions between poses.

 

The suggested reference was reviewed and added in line 444.

 

2. The confusion matrices for PSO and MicroGA appear identical. Please verify whether this reflects actual performance or a potential data entry error. Refer to the paper: Dynamic task allocations with Q-learning based particle swarm optimization for human-robot collaboration disassembly of electric vehicle battery recycling

 

Response

We thank the reviewer for this observation. In the revised version of the manuscript, the confusion matrices have been removed to avoid suggesting a classification-based evaluation, which is not the objective of this work.

 

Instead, the evaluation has been reformulated to focus on two complementary aspects: (i) the completeness of the generated transitions, assessed through the homotopy parameter $\lambda$, and (ii) the similarity between generated and reference sequences, measured using Dynamic Time Warping (DTW).

 

Specifically, we have included a table 4 reporting the fitness values associated with each generated animation, where values of $\lambda$ close to 1 indicate complete transitions. Subsequently, lines 383–395 present the similarity calculation using DTW. It can be observed that when the fitness of the individual is close to 1, the DTW distance is lower, indicating greater similarity between the generated action and the original one. In lines 396-409, in addition to reporting the mean and standard deviation of the DTW values, a 95\% confidence interval has been incorporated to better characterize the variability of the results, considering the stochastic nature of the optimization algorithms. In Figure 11, the impact of the standard deviation is explained in lines 363-369.

 

These changes provide a more appropriate evaluation framework aligned with the objectives of the proposed method, focusing on transition feasibility and sequence similarity rather than classification performance.

 

The suggested reference was reviewed and added in line 481.

 

3. Why was the number of intermediate frames fixed at 30? Could varying this number per action or pose pair improve motion naturalness?

 

Response

We thank the reviewer for this observation. The number of intermediate frames was fixed at 30 to provide a uniform temporal discretization of the generated motion sequences. This choice ensures consistency across all experiments and enables a fair comparison between different optimization methods.

 

It is important to note that increasing the number of intermediate frames does not necessarily improve the feasibility of the generated transitions. Instead, it increases the temporal resolution of the trajectory, resulting in more gradual changes between poses. Conversely, a smaller number of frames leads to coarser transitions with more abrupt variations.

 

We agree that adapting the number of intermediate frames for each action or pose pair could provide additional flexibility. However, this would introduce an additional parameter into the optimization process, increasing its complexity. Therefore, a fixed value was adopted in this work as a design choice.

 

 

4. The method selects key poses via k-means. How sensitive are results to k-means initialization? Were multiple runs performed to ensure stability?

Response

 

We thank the reviewer for this important observation. It is well known that the k-means algorithm is sensitive to initialization, which may lead to variations in the selected cluster centers. In this work, k-means is used as a preprocessing step to identify representative key poses from the input sequence. To reduce sensitivity to initialization, multiple initializations were considered, and the solution with the lowest within-cluster variance was selected.

 

It is important to note that the role of k-means in the proposed framework is to provide an initial set of representative poses, rather than to directly determine the final motion generation. The subsequent homotopy-based transition generation and optimization process further refine the resulting motion, reducing the impact of variations in the initial clustering.

 

Therefore, while some variability may arise from k-means initialization, its effect on the final generated sequences is limited within the context of the overall framework.

 

5. Please clarify how the homotopy approach handles occluded or missing joints in skeleton data, and whether joint correlations are preserved during transitions. Refer to the paper: Multi-scenario digital twin-driven human-robot collaboration multi-task disassembly process planning based on dynamic time petri-net and heterogeneous multi-agent double deep Q-learning network

Response

 

We thank the reviewer for this important question.

 

In the present work, the selected sequences from the dataset do not contain missing or occluded joint data. Therefore, the proposed framework operates under the assumption that complete skeleton information is available for all frames.

 

The homotopy-based approach generates transitions by continuously transforming the full set of joint positions from an initial pose to a final pose. As a result, the relative configuration of joints is preserved throughout the transition, maintaining the structural relationships defined by the input poses.

 

However, the method does not explicitly incorporate mechanisms to handle missing or occluded joints. In scenarios where incomplete skeleton data is present, a preprocessing stage would be required to reconstruct or estimate the missing joint positions before applying the proposed framework.

 

This limitation has been acknowledged in the revised manuscript in lines 90-92, and incorporating robustness to incomplete or noisy skeleton data is considered an important direction for future work.

 

The suggested reference was reviewed and added in line 481.

 

A general revision of the document was carried out, and the discussion and conclusion sections were updated.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

No further comments.

Author Response

We sincerely thank the reviewer for the valuable comments. We have carefully revised the entire manuscript to improve its clarity and writing quality. The tables and figures have also been thoroughly reviewed and enhanced accordingly. We hope that these revisions adequately address the reviewer’s concerns and meet the expected standards.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors substantially revised the manuscript in response to the previous review. The role of the homotopy parameter lambda is now clearly defined as an indicator of transition completion rather than a measure of perceptual realism. The manuscript explicitly separates the optimization objective (based on lambda) from the evaluation stage (based on DTW), resolving a conceptual inconsistency in the original version.

In addition, the authors moderated their claims, removed potentially misleading evaluation elements such as confusion matrices, and clarified the scope of the work as a feasibility study rather than a state-of-the-art motion generation method. Overall, the revised version is significantly clearer, more consistent, and better aligned with its actual contributions.

Despite the improvements, the experimental evaluation remains limited. The study does not assess generalization across multiple sequences, subjects, or noise conditions. Furthermore, the lack of perceptual evaluation or biomechanical validation means that the realism of the generated motions is not directly assessed. In addition, there is no comparison with state-of-the-art methods, and the statistical analysis remains relatively basic. These limitations restrict the impact of the work but are acceptable given the current approach, i.e. treating the work as a feasibility study.

With some hesitation, I would change my verdict to "accept with minor revisions". The paper still has notable problems, but at least the authors put in the effort to improve the mansucript, given the strict and limited deadline.

Several things that I would change before publication:

- Clarify early in the manuscript that DTW does not measure perceptual realism or physical plausibility.

- Extend and improve the discussion of limitations, particularly regarding the absence of perceptual and biomechanical evaluation.

- Provide a brief interpretation of the wide confidence intervals reported for DTW values.

- Perform minor language and grammar corrections throughout the manuscript, mostly to make the editors' job easier.

Author Response

Comment 1: Clarify early in the manuscript that DTW does not measure perceptual realism or physical plausibility.

Response:

Thank you for this important observation. We have clarified this point in the manuscript by adding explicit statements in the Abstract (lines 14–16) and the Introduction (lines 140–144). We now clearly state that DTW is used as a sequence similarity measure and does not evaluate perceptual realism or biomechanical plausibility. This clarification complements the explanation already provided in the Methodology section.

Comment 2: Extend and improve the discussion of limitations, particularly regarding the absence of perceptual and biomechanical evaluation.

Response:

Thank you for this valuable suggestion. We have expanded the discussion of limitations in Section 4 (Discussion) (lines 473–487) to provide a more detailed analysis of the absence of perceptual and biomechanical evaluation. We now explicitly discuss the implications of this limitation, including the lack of joint constraints, balance considerations, and physical plausibility, as well as its impact on complex motion generation. Additionally, we highlight this aspect as an important direction for future work.

Comment 3: Provide a brief interpretation of the wide confidence intervals reported for DTW values.

Response:

Thank you for this helpful suggestion. We have added a brief interpretation of the confidence intervals in the Experiments and Results section (lines 412–430 and 435–440). The revised text explains that the observed variability is related to differences in motion complexity across actions, rather than to instability of the proposed method

Comment 4: Perform minor language and grammar corrections throughout the manuscript, mostly to make the editors' job easier.

Response:

We sincerely thank the reviewer for the valuable comments. We have carefully revised the entire manuscript to improve its clarity and writing quality. The tables and figures have also been thoroughly reviewed and enhanced accordingly. We hope that these revisions adequately address the reviewer’s concerns and meet the expected standards.

Reviewer 3 Report

Comments and Suggestions for Authors

It can be accepted now for the responses and improvements.

Author Response

We sincerely thank the reviewer for the valuable comments. We have carefully revised the entire manuscript to improve its clarity and writing quality. The tables and figures have also been thoroughly reviewed and enhanced accordingly. We hope that these revisions adequately address the reviewer’s concerns and meet the expected standards.