Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The research addresses the challenge of achieving safe and stable trajectory tracking for quadrotors operating in environments with obstacles and model uncertainty. Specifically, it improves handcrafted dynamic models by using a data-driven approach.

The learning-based control is highly relevant. However, the originality is moderate. While the application to quadrotors is useful, the core components (HDM and CLF-ECBF-QP) are already established in prior literature. This submission primarily focus on integration, instead of the development of new frameworks.

It provides a implementation framework that combines physically consistent learning (Hamiltonian Neural ODEs) with safety guarantees (ECBF-QP) for quadrotors. It adds a specialized parameterization for the inertia matrix and a scaling proposition (Proposition 1) to address numerical conditioning in the optimization problem.

The conclusions are only partially consistent with the evidence. While the results show the framework outperforms a broken nominal model, they do not definitively prove that the Hamiltonian structure is the cause of this improvement, as there is no comparison with a standard learning baseline, a non-Hamiltonian neural network.

Most references are appropriate, but Reference 1 appears to be a placeholder template.

The major components of this paper, including the HDM and CLF-ECBF-QP control framework, are already established in previous literature. The real novelty is primarily to integrate these components and design specilized parameterization for quadrotor. The authors should be more explicit about this integrative work in their contribution statement, and clearly explain their work relative to prior Hamiltonian NN and CLF/CBF frameworks.

Meanwhile, the experiments do not directly verify that using a Hamiltonian NN is beneficial. The only comparisons are based on deliberately mis-specified nominal models. It is unclear that whether the Hamiltonian NN improves over simpler non-Hamiltonian NN.

Other minor issues:

1. The symbol distinguishing of rotation matrix and real number set should be more visible and explicit in Section 2.

2. What is \widetilde{f} in Equation 10?

3. The first reference seems to be a mistake.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1. Line #309: small rectangular object?
2. Experimental research does not include. It is strongly recommended to include any comparable test, if there is no authentic test during the research.
3. References: First reference does not clear what represents.
4. Please improve quality of figures.
5. Conclusion consists of just a few sentences. Readers are the most interested in concluded remarks and especially in propositions for the future works. Please provide wider concluded remarks.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Reviewer Report

Manuscript Title:
A Hamiltonian Neural Differential Dynamics Model and Control Framework for Autonomous Obstacle Avoidance in Quadrotor Subject to Model Uncertainty

General Comments

The manuscript presents a physics-informed, data-driven modeling and control framework for quadrotor systems based on a Hamiltonian Neural Differential Model (HDM) formulated on the SE (3) manifold. By explicitly studying inertia parameters and approximating the control input matrix using neural networks, the method addresses modeling uncertainty while preserving physical structure. The learned dynamics are reformulated into a control-affine form and integrated with a CLF–ECBF–QP controller to ensure stability and safety in obstacle-rich environments.

The topic fits well within the scope of Drones Jornal. The methodology is technically valuable, the formulation is mathematically consistent, and the simulation results support the claims. The manuscript is generally clear and well organized. Minor revisions are required.

Originality and Novelty

The integration of Hamiltonian learning on SE (3) with safety-critical CLF–ECBF–QP control is novel. The explicit learning of inertia parameters within a Hamiltonian framework provides improved interpretability compared to purely black-box models.

Scientific Soundness

The assumptions and modeling choices are reasonable and well justified. The Hamiltonian structure improves robustness and physical consistency. Simulation results demonstrate improved tracking accuracy and obstacle avoidance compared to nominal models with incorrect inertia or aerodynamic uncertainty.

Clarity of Presentation

The manuscript is clearly structured, and most figures are informative. Some sections would benefit from additional explanation.

Specific Comments

1. Introduction:
   Clarify the advantages of the proposed HDM compared to purely neural-network-based models. A brief quantitative comparison with a black-box baseline would strengthen the contribution.
2. Section 3.3 (HDM):
   It will be better if you provide additional details on the neural network architecture (layers, neurons, activation functions) and training parameters (optimizer, learning rate, number of epochs).
3. Section 4.3 (QP Controller):
   Include more intuition on the tuning of controller gains (γ₁, γ₂) and the inertia scaling factor (K = 100). The implications of assuming should be discussed.
4. Section 5 (Simulation Results):
   Consider adding an ablation study comparing the HDM with a neural-network-only model. A table summarizing RMS tracking errors and minimum obstacle distances for Figures 5 and 6 would strengthen the quantitative evaluation. Visualization of the learned Hamiltonian energy evolution would further support physical consistency claims.
5. Figures and Tables:
   Figure 4 should be introduced more clearly and detailed  in the text. Table 1 may be moved earlier for improved flow.
6. Practical Considerations:
   A brief discussion on sensor noise, computational cost, and possible extensions to multi-quadrotor systems would enhance practical relevance.

Recommendation

Minor Revision

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have address most of my concerns. The manuscript in the current version is good to go.