A Novel Approach for Self-Driving Vehicle Longitudinal and Lateral Path-Following Control Using the Road Geometry Perception

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposed a path tracking method considered road geometry factors, and the core of the proposed method is the DRL. However, the author did not give the detail on how to train the DRL model, and the training time, loss and other related factors were also not presented.

1.The author say it was a novel deep reinforcement learning for path tracking system, so where is the “novel” point ? considering the road geometry perception is the novel point? What was the difference between the novel DRL and DRL ? please explain it.
2.Section 4 “Materials and methods”, where is the materials? 
3.Section 6 “conclusions”, the conclusion means the authors summarized the main idea of the paper,  such as ,what methods was proposed, what results were obtained and what conclusions would be got from the results. The “conclusion” was not to summarize other researchers works, so why cited so many references in this section?
4.In table 1, typing errors, units in the first two lines, the “2” should be upper type, as presented in the third line of this table.
5.In fig. 1, how to get the actual position and velocity of the cars? What kind of data sets were used for training? Public database?  From the fig.2, the data seemed generated by simulation. 
6.In line 511, the results in the fig.3, the bottom figure, the jerk changes frequently, and not close to zero, why?
7.In line 376, Jk(t−1) should be jk(t-1), the parameter should be the same as presented in Eq.(11).

Author Response

Reviewer 1

Comments and Suggestions for Authors

This paper proposed a path tracking method considered road geometry factors, and the core of the proposed method is the DRL. However, the author did not give the detail on how to train the DRL model, and the training time, loss and other related factors were also not presented.

Thank you for the evaluation of our work. You are right regarding this comment, and accordingly, a more detailed description of the application of the DRL method has been added in the revised manuscript, subsection 4.2.

 

1.The author say it was a novel deep reinforcement learning for path tracking system, so where is the “novel” point? considering the road geometry perception is the novel point? What was the difference between the novel DRL and DRL? please explain it.

Thank you for this comment that helps us to better explain the main contribution of our work. You were right and the contribution is not on the technique used, DRL, but about the inclusion of a variable that is not usually considered and may make the driving safer. In this sense, this innovation comes from the fact that we did not find in the scientific literature other systems that introduce variables that sense and give meaning to what passengers may feel when travelling in a car. The novelty of this work regarding this point has been now highlighted in the introduction and conclusions sections. Even more, to clarify it the title of the paper has been changed.

 

2.Section 4 “Materials and methods”, where is the materials? 

Thank you. You are right. The title of section 4 has been changed to ‘Methodology’ as it better reflects the content.

 

3.Section 6 “conclusions”, the conclusion means the authors summarized the main idea of the paper,  such as ,what methods was proposed, what results were obtained and what conclusions would be got from the results. The “conclusion” was not to summarize other researchers works, so why cited so many references in this section?

Thank you for this suggestion. It was intended to provide a comparative approach of our proposal with others found in the scientific literature. But you are right and this summary of other research works is now included in the 'Results and discussion' section. The 'Conclusions' (section 6) have been restructured according to your indications.

 

4.In table 1, typing errors, units in the first two lines, the “2” should be upper type, as presented in the third line of this table.

Right. Thank you. Typos have been corrected.

 

5.In fig. 1, how to get the actual position and velocity of the cars? What kind of data sets were used for training? Public database?  From the fig.2, the data seemed generated by simulation. 

Thank you for this comment. In this first study, the environment is simulated. We have simulate different driving profiles (Urban Driving Cycle, UDC) that are used to simulate on-road testing of light commercial vehicles. Thus, no public databases are used, and a generic standard vehicle is parametrised.

 

6.In line 511, the results in the fig.3, the bottom figure, the jerk changes frequently, and not close to zero, why?

Thank you for this comment. The change is due to small speed variations, less than 1 km/h, which do not affect passenger comfort (Gillespie,1992).

 

7.In line 376, Jk(t−1) should be jk(t-1), the parameter should be the same as presented in Eq.(11).

Thank you. Corrected.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper analyzes the impact of the road geometry on vehicle path following, and proposes a novel RL-based framework for longitudinal and lateral vehicle control. 
The paper can be interesting, but it still needs to be improved in the following aspects.

Comments:
- From the abstract and introduction, it is not fully clear what the proposed system does and how it operates. The authors say that their DDPG agent computes the longitudinal controls, while a DQN does lateral control. However, it is not clear whether there is a reference trajectory to be tracked by these controllers. In other words, what is the purpose of the new controllers of this paper? Tracking a reference trajectory for path following? Please be more clear, both in the abstract and introduction.
- In the introduction, consider outlining the paper's contributions in the form of a list, to make them more clear.
- In the literature review, the authors could mention some other recent works about lateral vehicle control and path tracking using supervised learning and neural networks, such as "A physics-driven artificial agent for online time-optimal vehicle motion planning and control".
- After the introduction and related work analysis, it would be useful to have a section or subsection to present the framework with a high-level overview of the proposed system, possibly with a simple diagram to describe the framework. Otherwise, the high-level goal and methology of the paper are not clear.
- line 166: "design speed" --> do you mean the desired (or reference) speed? And why is it computed with the equation (1)? What advantage does (1) bring in comparison with other possible designs? Please clarify this.
- line 168: "f_t is the maximum coefficient of cross friction" --> what do you mean by cross friction? Is it the friction coefficient between the tires and the road?
- line 180: f_t is redefined as "the maximum transverse friction mobilized coefficient". Please do not redefine variables in the text. If needed, introduce a new variable name. Moreover, "friction mobilized coefficient" is a very confusing term. Please clarify this and if possible use a more standard term.
- line 183: a_road is not in math mode
- line 185: v_l and not vl
- there are a lot of typos like the ones above in the paper. Please fix them. I cannot list them all here. Another example is the units of measurement in Table 1 --> m/s^2 in the first two rows
- In (3), a_road seems to be in the longitudinal direction (due to the road slope), while in (4) it seems to be in the lateral direction. Please clarify.
- line 186: "perceived lateral acceleration caused by risky driving behaviour " --> this is very confusing and unclear. What is a risky driving behavior? 
- the jerk in eq. 6 does not need to be defined in a scientific paper of this type. It is a well-known concept.
- title of section 4.1 --> "to help sustainability" --> this is very generic. This paper does not seem to be about sustainability. Sustainablity should be mentioned only when discussing the possible impacts of the proposed system.
- In the results section 5, the storyline needs to be improved. Please do not use a lot of lists to simply describe what each figure shows, but be more narrative and critical in analyzing the results and motivate them. 
- The steering angle in fig. 4 is very jerky and oscillating. Would it be feasible for a real car? Please discuss this.
- Is there noise in your vehicle simulation environment? 

Author Response

Reviewer 2

Comments:

- From the abstract and introduction, it is not fully clear what the proposed system does and how it operates. The authors say that their DDPG agent computes the longitudinal controls, while a DQN does lateral control. However, it is not clear whether there is a reference trajectory to be tracked by these controllers. In other words, what is the purpose of the new controllers of this paper? Tracking a reference trajectory for path following? Please be more clear, both in the abstract and introduction.

 

Thank you for this comment. You are right, and we may have failed to explain this point clearly. It is a path-following control. The reference is the lead car. Therefore, there is no predefined trajectory, and in this sense, it is not a trajectory tracking. This has been clarified in the manuscript.

 

1- In the introduction, consider outlining the paper's contributions in the form of a list, to make them more clear.

Thank you for your suggestion. The main contributions of this study have been outlined in the introduction as a list to make them clearer.

2- In the literature review, the authors could mention some other recent works about lateral vehicle control and path tracking using supervised learning and neural networks, such as "A physics-driven artificial agent for online time-optimal vehicle motion planning and control".

Thank you for your help to improve the state of the art. This reference has been added among others.

3- After the introduction and related work analysis, it would be useful to have a section or subsection to present the framework with a high-level overview of the proposed system, possibly with a simple diagram to describe the framework. Otherwise, the high-level goal and methodology of the paper are not clear.

Thank you for this suggestion. A subsection where vehicle control system is introduced is now added.

 

4- line 166: "design speed" --> do you mean the desired (or reference) speed? And why is it computed with the equation (1)? What advantage does (1) bring in comparison with other possible designs? Please clarify this.

Thank you for this comment. The design speed is defined as the maximum speed that can be safely maintained on a given section of a road with low traffic volume and in good weather conditions. Under those conditions, the design speed is determined solely by the characteristics of the road.

It is applied in all countries and is used as a reference for the circulation of a vehicle over a stretch of road with a given curvature in safe conditions. The following regulations can be consulted:

Highway Capacity Manual 7th Edition: A Guide for Multimodal Mobility Analysis. Washington, DC: The National Academies Press. National Academies of Sciences, Engineering, and Medicine. 2022. https://doi.org/10.17226/26432

 

5- line 168: "f_t is the maximum coefficient of cross friction" --> what do you mean by cross friction? Is it the friction coefficient between the tires and the road?

Thank for your question. Yes, the coefficient of cross friction (or lateral friction coefficient, maximum transverse friction mobilized coefficient) refers to the friction between the road surface and the tires, specifically in the lateral direction. It depends on factors like road surface, tire composition, and weather conditions. This explanation has been added to the manuscript.

 

6- line 180: f_t is redefined as "the maximum transverse friction mobilized coefficient". Please do not redefine variables in the text. If needed, introduce a new variable name. Moreover, "friction mobilized coefficient" is a very confusing term. Please clarify this and if possible use a more standard term.

Thank you. You are right. We refer to the previously cross friction coefficient. The text has been revised for consistent concept definitions.

 

7- line 183: a_road is not in math mode

Thank you. Typo has been corrected.

 

8- line 185: v_l and not vl

Thank you. Corrected.

 

9- there are a lot of typos like the ones above in the paper. Please fix them. I cannot list them all here. Another example is the units of measurement in Table 1 --> m/s^2 in the first two rows

You are right. Thank you. The manuscript has been thoroughly revised to correct the typos.

 

10- In (3), a_road seems to be in the longitudinal direction (due to the road slope), while in (4) it seems to be in the lateral direction. Please clarify.

Thank you for your questions. Acceleration due to road refers to lateral acceleration caused by the road's cross slope or camber (transversal gradient), affecting the vehicle body and passengers. On the other hand, road slope is inclination of road horizontally in the forward direction of the road. These definitions have been added to the manuscript.

 

11- line 186: "perceived lateral acceleration caused by risky driving behaviour " --> this is very confusing and unclear. What is a risky driving behavior? 

Thank you for this comment. Risky driving behaviour includes actions that increase accident chances, endanger people, and reduce road safety. Common causes are poor judgment, distraction, speeding, and drunk driving. Therefore, by measuring the sensation of lateral acceleration, we can deduce when inappropriate driving is occurring. For clarification and improve the quality of the explanation this sentence has been rephrased by “can be caused by not suitable driving behaviour or abrupt road curvature”. The following regulations can be consulted:

National Highway Traffic Safety Administration (NHTSA). (2021). Traffic Safety Facts: Risky Driving. Retrieved from https://www.nhtsa.gov

 

12- the jerk in eq. 6 does not need to be defined in a scientific paper of this type. It is a well-known concept.

Thank you for the comment. Definition has been removed.

 

13- title of section 4.1 --> "to help sustainability" --> this is very generic. This paper does not seem to be about sustainability. Sustainability should be mentioned only when discussing the possible impacts of the proposed system.

Thank you for your comment. Title of section 4.1 has been changed for “A vehicle system to assist in the driving task.”

 

14- In the results section 5, the storyline needs to be improved. Please do not use a lot of lists to simply describe what each figure shows, but be more narrative and critical in analyzing the results and motivate them.

Thank you for your comment. Section 5 has been re-structure to include a more critical analysis of the results.

 

15- The steering angle in fig. 4 is very jerky and oscillating. Would it be feasible for a real car? Please discuss this.

 

Thank for this interesting comment. This may be because the system is trained with a sinusoidal signal, as is commonly done in other studies (Gillespie, 1992). For smooth and controlled driving, the recommended steering angular velocity is ≤ 0.5 rad/s. In our case, it is always less than 0.4 rad/s in all experiments. It is usually around 0.2 rad/s except in some sections with curves or speed changes from the preceding car. For these reasons, we believe it would be feasible for a car, although a possible future action would be to try to smooth the control to reduce these oscillations.

 

16- Is there noise in your vehicle simulation environment? 

 

Thank you. The vehicle simulation environment is noise-free. To simulate a more realistic scenario, noise could be incorporated, but in any case, it should be filtered out at a later stage before working with the signals.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have revised all according to the previous comments. 

there are still some parts need to revise:

1. the formatof the references should be revised. The published years of most of the references was added after author list, such as in reference 5, 6, but the 7,8 9,10, the year was listed in the end. In addition, some journal name was wrote by italic, and some were not . Please check all the references and using the same format.
2. Reference 20 , 28 , 30, no page range and volume.
3. Reference 37, page, year, journal name.
4. Reference 38, two references were mixed in one.
5. Reference 39, need to check the format.

Author Response

Reviewer 1

Comments:

- there are still some parts need to revise:

1. the formatof the references should be revised. The published years of most of the references was added after author list, such as in reference 5, 6, but the 7,8 9,10, the year was listed in the end. In addition, some journal name was wrote by italic, and some were not . Please check all the references and using the same format.
2. Reference 20 , 28 , 30, no page range and volume.
3. Reference 37, page, year, journal name.
4. Reference 38, two references were mixed in one.
5. Reference 39, need to check the format.

You are right. Thank you for appreciating. All references have been revised and corrected.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The changes made by the authors are appreciated, and the paper can now be considered ready for publication.

Author Response

Reviewer 2

Comments:

- The changes made by the authors are appreciated, and the paper can now be considered ready for publication.

 

Thank you so much. I am very grateful for the appreciation of our work. Your acknowledgment serves as a great motivational boost and reinforces our commitment to delivering the best results possible

Author Response File: Author Response.pdf