Vehicle Safety Planning Control Method Based on Variable Gauss Safety Field

Round 1

Reviewer 1 Report

Vehicle planning control is a topic on which many papers and books have been published. Of course, some contributions are possible because of the difficulty of the topic but some practical validation would be advisable. For these reasons, the following ideas must be considered to improve the paper:

- The are too much work done previously in the literature tan the state of the art presented in the paper. Please, complete it and look for some review papers and books that deal with the same topic.

- Contributions are not completely clear to be really novel and significant. Plese, clarify them to show novelty in comparison with other research.

- The theoretical approach is quite convectional. Highlight differences with other approaches

- Results are based on simulations on CARLA. Is this representative enough and can these results be considered are representative of real data? Is there any validation? Decision making in vehicles is really a challenging problem in real life with unexpected scenarios. Is this fact considered in simulations?

- I suggest changing the title of the last section. There is no point to name it "Results and Conclusions". Conclusions should be enough.

Author Response

1. Supplement the variable Gaussian field part of the overview. It also complements the advantages of others' literature.

2. Redescribe the contribution of this article.

3. Figure 14 shows the comparison experiment between the safe autopilot trajectory planning method based on the time series bird's-eye view and the policy gradient algorithm used in this paper and other methods. The experimental results show that the method used in this paper improves the convergence efficiency of the algorithm, and is basically the same as the DDPG algorithm in terms of average reward. It reflects the advantages of this method.

4. At present, most of the reinforcement learning is done based on the simulator. It is a disadvantage of this paper that unexpected situations are not considered in this paper. We will discuss them in the final summary and outlook.

5. Change the last section title as ‘Conclusions’.

Reviewer 2 Report

In this paper, the authors propose a Gaussian field-type model of a vehicular traffic scenario for automatic safe navigation planning, considering that all of the environment and vehicle characteristics and variables are known.

The authors mention that the bird's eye view is built with a perception module of the automobile, but this is not a trivial task and cannot be omitted in the paper. In the opinion of this reviewer, this is the weakest point of the proposal because this scenario reconstruction could be slow and fair complicated per see.

Even more, the proposal must be compared with other similar approaches; indeed, the literature review must be improved since there are many similar proposals, for instance, Biological Action-Perception, and Attractor-Dynamic, among others.

A grammar review must be performed on the entire paper; there are mistakes.

Authors must improve Figure 2 and its explanation; a non-specialized reader could find this figure hard to understand.

The white lines in Figure 3 are not legible.

Line 27, page 4, the notation must refer to a single dot variable as speed and a double dot for acceleration; the jerk is a triple dot; please check.

Authors must ensure that all figures are completely legible; actually, almost all are of low quality, degrading the paper's quality. Even more, authors must make an effort to describe wholly and correctly all the figures. 

Line 8, page 6, what Gaussian function? which sigma?

The enumeration of equations is repeated, which is unacceptable.

Please explain the terms of the reward function; for instance, what additional coupling reward is?

How that k1,...,k6 (and other constants) values selected? (justify)

Author Response

1. The gray line in Figure 3 has been modified to become clearer.
2. In line 22 of page5, the variable corresponding to the symbol has been modified.
3. The clarity of Fig. 3 13 14 15 12 has been improved because it is difficult to find the data with long time.
4. Formula (1) and Formula (4) are not enumerated repeatedly, and the parameters are different. One is static and the other is dynamic.
5. The function of reward function has been explained.
6. The value comparison diagram of K1-K6 has been given.
7. Figure 14 shows the comparison experiment between the safe autopilot trajectory planning method based on the time series bird's-eye view and the policy gradient algorithm used in this paper and other methods. The experimental results show that the method used in this paper improves the convergence efficiency of the algorithm, and is basically the same as the DDPG algorithm in terms of average reward. It reflects the advantages of this method.
8. The explanation for the time sequence aerial view of Figure 2 has been given.

9. Supplement the variable Gaussian field part of the overview. It also complements the advantages of others' literature.

10.  Explanations on dealing with different complex scenarios and vehicle decisions are made in the last part of the article.

Reviewer 3 Report

This paper proposes an automatic driving trajectory planning method based on a variable Gaussian safety field. The paper is well written and organized. The following suggestions are included in the revised paper.

(1) Compared with the proposed method, how about the advantages of the others?

(2) Except for the highway scene, how about the proposed method used in other cases?

(3) Why could the experimental results be carried out to further to verify the advantages of the proposed method?

Author Response

1. Figure 14 shows the comparison experiment between the safe autopilot trajectory planning method based on the time series bird's-eye view and the policy gradient algorithm used in this paper and other methods. The experimental results show that the method used in this paper improves the convergence efficiency of the algorithm, and is basically the same as the DDPG algorithm in terms of average reward. It reflects the advantages of this method.

2. Explanations on dealing with different complex scenarios and vehicle decisions are made in the last part of the article.

3. At present, most of the reinforcement learning is done based on the simulator. It is a disadvantage of this paper that unexpected situations are not considered in this paper. We will discuss them in the final summary and outlook.

Round 2

Reviewer 1 Report

My previous comments have been considered. But the last new paragraph in conclusions section is not completely appropriate. It seems as future works that will not be done. I suggest a reformulation discussing limitations of the paper and how these future works would improve research.

Author Response

The future outlook has been modified, please check.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors have corrected most of the observed. 

However, the ENUMERATION of equations is still repeated, which is unacceptable; please authors, read carefully.

Also, authors must include how in a real scenario, such bird's eyesight is obtained from sensor data and how a non-exact build affects the algorithm performance in a real scenario.

Author Response

Dear reviewer, this paper mainly focuses on the subsequent planning control part. It does not involve the generation of the aerial view, which is generated by the CARLA simulator and has no relationship with the sensor.

The ENUMERATION of equations is solved.

Author Response File: Author Response.docx

Reviewer 3 Report

Authors have responded all my questions. Thank you.

Author Response

Thank you for your suggestions very much!