Efficient Focus Autoencoders for Fast Autonomous Flight in Intricate Wild Scenarios

Round 1

Reviewer 1 Report

An autonomous navigation method is proposed named SAC_FAE. The topic is interesting, but there are still large rooms to improve the quality of the manuscript.

1.     What information obtained from the onboard inertial measurements has been used for the proposed navigation method.

2.     How to integrate the inertial measurement information and visual information?

3.     The realization steps of the proposed method should be explained in more details. How it is used in flying robot hardware.

4.     How fast the robot flies? Is it can meet the requirement of “Fast Autonomous Flight” as the title of manuscript said?

5.     Is the proposed method tested on the real scenarios as described in Fig. 5? If so, more details about the hard-ware of the flying robot should be explained.  

6.     In “Conclusions” part, the author claims that the “To demonstrate the effectiveness of our strategy, we conducted multiple experiments ranging from simple to complex scenarios in urban and forest environments.” But, current manuscript hasn't presented a substantial amount of experimental results.

English is fine

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

1. Could you please explain the reason behind the rate of change in the heading angle of the agent being restricted to the range of [-90.0, 90]rad/s in section 3.2.2 of the experiments?

2. Table 1 displays various hyperparameters used during the policy training process. Could you elaborate on the reasons behind setting these values and what they represent?

3. In section 5 Results, the comparison of different algorithms is discussed. Could you provide more information on the parameters and assumptions utilized by these algorithms?

4. It is suggested that an algorithm assumption be included. What is the reason behind this recommendation?

5. It is recommended to include an algorithm flow chart and demonstrate the effectiveness of the algorithm with a simple example. Could you please provide this information?

6. Could you explain the differences and connections between the scene simulation environments and why these scenes were chosen for testing?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors introduced the SAC_FAE algorithm in this study for achieving high-speed navigation in complex environments through the use of deep reinforcement learning (DRL) policies. Their methodology incorporated a soft actor-critic (SAC) algorithm along with a focus autoencoder (FAE). The authors' end-to-end DRL navigation approach empowered a flying robot to efficiently carry out navigation tasks without relying on pre-existing map information. Instead, it relied solely on the front-end depth frames and the robot's own pose information. The authors' algorithm surpassed the performance of existing trajectory-based optimization methods, showcasing its efficacy and efficiency in various testing environments with flight speeds surpassing 3 m/s. This underscores its ability to enable rapid navigation through intricate landscapes, fundamentally reshaping the landscape of autonomous aerial exploration. The paper is clear, however needs improvements:

1. Introduction needs improvement and more recent studies must be added.

2. Highlight the gaps left by the previous approaches.

3. Please compare the proposed framework with other method.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The presented paper is enriched with the proposed method description. The authors present a large number of schemes, graphs and algorithms supporting the presented method's superiority over the traditional ones.
However, the reviewer has a number of remarks:
1. The results should be described in an impersonal manner.
2. The advantages of specifically fast-flight applications are not clear. Is it possible to perform mapping tasks in such a case? What are the benefits of fast flight?
3. The proposed algorithm's operational speed is not enlightened. While quality can be judged from the presented results, the timing characteristics are not.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Accept in present form

Reviewer 4 Report

I thank the authors for their detailed answers. I am completely satisfied with the answers. The new version of the paper removes all previously raised questions.

I recommend the paper for publication.