GPU-Accelerated Multi-Objective Optimal Planning in Stochastic Dynamic Environments
Round 1
Reviewer 1 Report
The authors proppsed a GPU accelerated path planning algorithm for unmanned surface vehicles (USV) under scenarios with dynamic obstacles, and the effect of uncertain surface water currents. Multiple objectives (travel time and energy/net energy consumption minimisation) are optimised in path planning from known start location to destination. The presented work is an extension of the authors' work on path planning with single objective. Simulations of two mission scenarios were conducted to show the effectiveness of the proposed method.
The simulator involved in the study is over-simplified. The simulation domain (spatial-temporal state space) was discretised into a grid of 100x100x120 (each spatial cell corresponds to a 5 km x 5 km region and the time step involved is 1.728 hr) and so was the action space. The USV is considered as massless point in simulation. How realistic is the simulation model compared to real underwater scenarios? What is the target velocity of the USV under consideration? Please justify the validity of the simulator settings.
How are probability distributions or samples (w_v and w_g respectively) of the dynamic vector fo surface current and scalar field of harvested energy obtained/estimated from the real environment? When the proposed algorithm is applied in real missions, what sensing systems are needed to take measurements of w_v and w_g for planning?
The authors are suggested that they discuss how well the scalarisation by weighted combination method in handling multiple conflicting objectives. Have you considered other schemes to handle multiple objectives?
The authors are suggested that they compare the performance of the proposed algorithm with other multiobjective optimisation algorithms, NSGA II.
The authors are suggested that they evaluate the performnance of the proposed algorithm using a sophisticated physics based simulation for underwater robotics which model realistic hydrodynamic effect and dynamics of the USV.
The authors are suggested that they demonstrate the practicality of the proposed method by conducting experiments with physical USV.
The authors are suggested that they discuss the extensibility of the proposed method to handle multi-robot path planning and/or formation control scenarios.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The result shown in Figure 4 and Figure 5 needs a detailed explanation for the reader to clearly understand. The reader is difficult to relate the diagram with the explanation given.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
- General comments
(1) In this paper, the instantaneous path of the agent is directly determined by the vector sum of the velocity field and the agent’s action. In several situations, this assumption is appropriate. However, in other situations, this can be an oversimplification, so the path planner in this paper may not be necessarily adequate for practical applications.
(2) Many necessary details are not explained in this paper. Here are some examples. The ship domain, moving speed, and intervals of the obstacles are not described. How obstacles affect the reward function is not explained. In Fig. 4, the radiation field changed but the reason is not explained. In Figure 5B for p0 = 20, X = 70, and Y = 60 or so, the agent should have crossed the channel, but the path was still not reasonable, which was not explained. Besides, the flow field should also be summarized clearly because it is an important part of this paper.
(3) “GPU-Accelerated” is just a way to calculate data in a computer, not an innovation in research. If the authors want to make this one of the highlights of this paper, the focus should be not on the specific process of the algorithm, but on how much influence the GPU has on which part of the algorithm, the rationale for it, and how much it differs from the CPU-based algorithm. In addition, analyzing whether the proposed algorithm can provide the same acceleration for larger and more complex, and specialized scenarios is also one of the issues to be discussed.
(4) In the simulated scenarios in this paper, the agent can take actions of only 2 different speeds. According to this line of thinking, more classes of different speeds such as cruising speed, full speed, and zero speed should also be considered for practical applications. More importantly, most of the magnitude of the velocity field exceeds the maximum speed of the agent. This is true for low-speed AUVs. However, the maximum speed of some AUVs is over 3 m/s. Thus, in section 4, it's unreasonable to set the agent's maximum speed to only 70 cm/s and a higher speed places higher demands on the rationality and performance of the planner than the proposed scenarios.
(5) In section 4.2, the interval between the obstacles is about 250 km and the total navigation time is about 8 days, which is inconsistent with the scale of the collision avoidance mission. The so-called “shipping channel” in this paper is more like an ideal scenario for the low-speed agent rather than a simulated busy shipping channel with the agent under the head on, crossing, overtaking, or complex encounter situation. Although with stochasticity, the proposed double gyre flow environment with obstacles moving at long intervals in a line is neither large and complex enough for optimal routing test nor suitable for collision avoidance test.
(6) In a word, a single simple mission environment with only and finish point, a small number of actions the low-speed agent can take, and unsuitable collision avoidance scenario in this paper failed to prove the applicability of the proposed path planner’s ability to plan and optimize the path under the influence of a large range of stochastic dynamic flow and moving obstacles. There are no strong tests in this paper to prove that the proposed planner can use the realistic environment data and the algorithm works with any complex flow fields, yet it is described as such in the conclusion.
- Specific comments
(1) The symbols and figures are somewhat confusing. For example, the description of symbol p0 in Fig. 1 in section 1 appears in section 4.2 and the meaning of symbol d is never explained.
(2) Fig. 1: “v(x, t; ω)” -> “v(x, t; ωv)”, “a = π(w, t)” -> “a = π(x, t)”, “x0” -> “xs”
(3) The caption of the figure should be brief and the detailed description of the figure should be in the text.
(4) Equations in the text such as the equation R1, gen(s, a; s′) = -Δt in Line 136 should be numbered.
(5) It is inappropriate to use different expressions to define the same symbol Rα in Eq. 3 and Eq. 4.
(6) In Fig. 4 and Fig. 5, several different paths are shown in each map. Because the lines overlap, it is hard to check whether a specific path is reasonable.
(7) The language in this article should be re-edited and corrected (e.g. Line 71: “etc” and Line 87: “makes”).
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors have addressed my comments successfully and therefore I recommend acceptance the revised manuscript for publication after careful proofreading of the manuscript so that grammatical and spelling mistakes are identified and corrected.
Reviewer 3 Report
I confirm that the authors have corrected this manuscript and
included additional information/data based on all of my
comments and recommendations.
Therefore I don't have any further comments regarding the
revised paper.
