Simulation Training System for Parafoil Motion Controller Based on Actor–Critic RL Approach
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors present work on employing reinforced learning to the control of a parafoil. The method as such and the results displayed are interesting and plausible. Especially to use the simulation model for training and using then the result for controlling the real system is good. Some criticism can be made and changes are suggested. In general, there is a tendency to use many words to explain things which can increase the number of pages but not necessarily contributes to understand the paper.
In the last sentence of the Abstract the abbreviation PPO for Preferred Provider Organization is used without introducing it previously.
The method developed is compared to a system with a traditional PID controller. Not much is said about the PID controller. There are many potential PID controllers. Is it optimized in any sense? How does it work.
Also the detailed description of the method in the Appendices has some deficiencies. The description of the Action in Appendix B is very short. Since it is at the core of the work, it should be more detailed. Furthermore, a quantity Ar is mentioned that has not been defined before. Also after formula A2 there is a Fir and after formula A4 there is an Ap. What does it mean?
From the point of view of the referee the authors should comment on the points mentioned before publication.
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe topic considered in the paper is interesting, important, and fits well into the scope of the journal. However, there are some issues to be corrected.
1. It is necessary to describe the aerodynamics in much more details. How aerodynamic forces are calculated, what assumptions are made, which model for aerodynamic forces is used, which software is used for calculation (if any), etc.
2. It is also necessary to describe in more details the wind models used. What are characteristics of the stochastic process? What is the structure and characteristics of gusts? Just references are not sufficient.
3. The equations of motion should be discussed in more details. The Authors mention only the “second law of Newton”. Does this mean that the rotation of the system with respect to the center of mass is neglected? If yes, then the general mechanical model is inadequate. If not, then something should be said about the corresponding equations describing the change in the spatial orientation of the system.
4. The control action should be described in more details. For instance, are accelerations (Fig.3) always the same? What is the output of the parafoil motion controller? Are these the required control forces, of the required speeds of control ropes? Maybe, the easiest way would be to provide the equations of motion of the system with corresponding comments.
5. The canopy is not, strictly speaking, a rigid body; therefore, something should be said about simulation of its dynamics.
6. The term “experiment” is used ambiguously. In fact, the paper discusses only one experiment (sec. 6, real flight). The other tests, strictly speaking, are not experiments, but numerical simulations.
7. The neural networks used in the paper should be described and commented in more details. In particular, clear information about inputs and outputs should be provided. The choice of network structure and activation functions should be discussed and some references should be given.
8. In 3.5, the Authors mention the “Euler angle”. Which angle is meant?
9. Numbers in Fig. 16a are difficult to read.
10. It is stated that the input layer of the MLP is 7×64. Is this correct? There are 17 inputs.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThe Authors improved the paper considerably, and I think that it can be now accepted for publication.
