An autonomous underwater vehicle path planning method based on terrain risk manifold and double-layer graph guidance

Abstract

This invention relates to the field of intelligent navigation and path planning technology for autonomous underwater vehicles (AUVs), and discloses an AUV path planning method based on terrain risk manifold and two-layer graph guidance. This method effectively solves the technical challenge of balancing path planning safety and computational real-time performance in complex seabed environments by constructing a risk manifold field that fuses multi-dimensional terrain features and introducing a two-layer dimensionality reduction search architecture. The risk manifold field provides global and continuous risk avoidance guidance for the vehicle, enabling the path to actively avoid high-risk areas such as steep slopes and extremely rugged terrain. The two-layer graph guidance mechanism precisely limits the high-resolution graph search overhead in large-scale environments to local safety corridors, greatly improving the computational efficiency and engineering deployment feasibility of the algorithm while strictly ensuring the dynamic safety baseline of the underwater vehicle.

Description

Technical Field

[0001] This invention relates to the field of intelligent navigation and path planning technology for autonomous underwater vehicles, specifically to a path planning method for autonomous underwater vehicles based on terrain risk manifold and two-layer graph guidance. Background Technology

[0002] Autonomous Underwater Vehicles (AUVs) are crucial equipment for exploring and developing marine resources, playing an irreplaceable role in fields such as marine mapping, resource exploration, underwater rescue, and military reconnaissance. In these missions, AUVs typically need to navigate extremely complex, unknown or semi-unknown seabed terrain, including seamounts, ridges, trenches, and deep-sea canyon walls. Therefore, how to plan an efficient, smooth, and dynamically constrained obstacle avoidance path in real time while ensuring absolute navigation safety has become a core technological challenge in the field of AUV intelligent navigation and control.

[0003] Existing AUV path planning techniques for complex terrain mostly rely on grid-based graph search algorithms, such as the traditional A* (A-Star), D* (D-Star), and Theta algorithms. However, these conventional single-layer grid search methods face significant technical bottlenecks when dealing with large-scale, high-resolution underwater environments with dramatic topographic relief. First, traditional single-layer shortest path strategies typically only use the physical length of the path or simple binary obstacles as optimization objectives and constraints. This strategy lacks sufficient sensitivity to complex terrain geometry features (such as steep slopes and extremely high surface roughness), easily leading the AUV to hug high-risk seabeds or traverse dangerous areas with drastic terrain changes, thus significantly increasing the risk of bottom contact or collision.

[0004] To compensate for the aforementioned safety deficiencies, some studies have proposed single-layer weighted search strategies that incorporate risk cost values. While these methods improve path safety by converting terrain hazard levels into costs, the number of nodes often increases exponentially when processing high-resolution, large-scale underwater raster maps. The enormous search space leads to a sharp increase in computational costs, often requiring solution times of several seconds or even tens of seconds, which is unacceptable for AUV online real-time planning systems with limited computing resources and a need for rapid response to sudden environmental threats.

[0005] Furthermore, to achieve smoother paths to accommodate the under-actuated navigation characteristics of AUVs, arbitrary angle planning methods (such as the Theta* algorithm) have been widely adopted. These methods break the limitation of traditional grid systems that only allow movement in fixed directions, and indeed achieve significant advantages in local path smoothness. However, arbitrary angle planning requires extremely frequent line-of-sight checks during node expansion. When dealing with large-scale, complex, undulating underwater terrain, this dense line-of-sight check mechanism incurs extremely high computational costs, resulting in very high overall time consumption and severely restricting its online deployment and real-time dynamic replanning capabilities in practical engineering applications.

[0006] In summary, existing path planning methods struggle to achieve an effective balance between path safety, real-time computation, and engineering deployability. Achieving highly interpretable terrain risk modeling in complex deep-sea environments, while strictly adhering to safe navigation constraints and significantly reducing the computational overhead of graph search in large-scale environments through algorithmic architecture innovation, has become a pressing technical challenge for those skilled in the art. Summary of the Invention

[0007] To address the technical problem that existing single-layer graph search algorithms struggle to balance path safety and real-time computation in complex underwater terrain, this invention provides an autonomous underwater vehicle path planning method based on terrain risk manifolds and two-layer graph guidance. The technical solution is as follows:

[0008] Step 1: Obtain the elevation grid data of the underwater topography, construct a basic environmental map matrix including X coordinates, Y coordinates and corresponding elevation Z, and adaptively estimate the physical step size of the grid by calculating the median absolute value of the spatial difference between adjacent grids in the corresponding dimensions of the X matrix and Y matrix respectively.

[0009] Step 2: Extract multi-dimensional terrain features based on the basic environment map matrix, and perform robust normalization on the multi-dimensional terrain features to construct a global terrain risk manifold field;

[0010] Step 3: Downsample the global terrain risk manifold field and construct a low-resolution coarse semantic map, and plan a global coarse skeleton path within the feasible region of the coarse semantic map;

[0011] Step 4: Map the global coarse skeleton path back and generate a dynamic corridor mask, and perform a refined path search within the feasible region defined by the dynamic corridor mask;

[0012] Step 5: Extract the corresponding three-dimensional coordinate sequence from the refined path points obtained by the search and use a sliding average filter with a preset window size to perform independent one-dimensional smoothing on each dimension of X, Y, and Z. Eliminate the jagged edges of the grid search by using an edge value copying filling strategy. Then perform dynamic and safety constraint verification along the smoothed path points.

[0013] Step 6: During the autonomous underwater vehicle's navigation along the planned path, gating replanning monitoring is implemented. By rolling the calculation of the incremental gradient of risks along the path within the forward-looking local window, the danger of unknown or abrupt terrain ahead is predicted in real time, and finally a planned path with dynamic adaptive capabilities is output.

[0014] Furthermore, when the median absolute value is not a finite value or is not a positive number, a preset default step size constant is set.

[0015] Furthermore, step 2 specifically includes the following steps:

[0016] Step 2.1: Calculate the spatial gradient of elevation Z in the x and y directions to obtain the local slope angle, as shown in the following formula:

[0017] ,

[0018] in, The local slope angle is robustly normalized to obtain the slope characteristics;

[0019] Step 2.2: Calculate the absolute value of the Laplace operator based on the second derivative of elevation Z, using the following formula:

[0020] ,

[0021] in, Let the absolute value of the Laplace operator be... As a proxy index for local roughness, and after robust normalization, roughness features are obtained;

[0022] Step 2.3: Calculate the absolute deviation of the local depth relative to the global median, and perform robust normalization to obtain the depth deviation feature;

[0023] Step 2.4: Generate a binary passable mask based on whether the local slope angle is less than the maximum allowable slope. Use a convolution kernel of a preset dimension to locally smooth the mask to calculate the local continuous passability ratio, and take its complementary value as the connectivity penalty feature.

[0024] Step 2.5: Perform linear weighted fusion of the slope feature, roughness feature, depth deviation feature, and connectivity penalty feature according to the set weights, and truncate to... For a closed interval, construct the global terrain risk manifold field as follows:

[0025] ,

[0026] Where S represents the slope feature, R represents the roughness feature, E represents the depth deviation feature, and C represents the connectivity penalty feature. For the global topographic risk manifold field, , , , Set weights for the corresponding features.

[0027] Furthermore, the robust normalization process specifically involves extracting the 1% to 99% quantiles of the terrain feature data in each dimension as upper and lower bounds for linear mapping, thereby eliminating the influence of extreme outliers.

[0028] Furthermore, step 3 specifically includes the following steps:

[0029] Step 3.1: Introduce a downsampling factor and use a block averaging strategy to downsample the global terrain risk manifold field, local slope angle, and local ruggedness proxy index to construct a low-resolution coarse-layer map.

[0030] Step 3.2: Within the feasible region of the low-resolution coarse-layer graph, a heuristic graph search algorithm is used to plan a global coarse skeleton path, where the node transition cost function is a composite cost function, designed as follows:

[0031] ,

[0032] Among them, 1.0 is the base step size cost. These are the downsampled global topographic risk manifold field, local slope angle, and normalized local ruggedness proxy index, respectively. Hyperparameters are penalized as a cost.

[0033] Furthermore, step 4 specifically includes the following steps:

[0034] Step 4.1: Map the global coarse skeleton path back to the original high-resolution fine-layer map, and perform densification processing between adjacent skeleton nodes using a linear interpolation algorithm.

[0035] Step 4.2: Extend the preset pixel radius along the densified skeleton nodes to generate a dynamic corridor mask.

[0036] Step 4.3: Within the intersection of the dynamic corridor mask and the original high-resolution feasible region, perform a fine graph search using the same composite cost function as the coarse-layer search to obtain fine-layer feasible paths.

[0037] Step 4.4: If no connected path can be found within the space defined by the dynamic corridor mask, a rollback mechanism is triggered, the corridor mask constraint is automatically discarded, and the constraint graph search is re-executed within the high-resolution feasible region of the entire graph.

[0038] Furthermore, the dynamics and safety constraint verification includes assessing whether the maximum slope exposure along the route exceeds the dynamics threshold, calculating the average and maximum risk limits of the path, and calculating the cumulative yaw cost of adjacent heading angles.

[0039] Furthermore, when the risk increment gradient exceeds the set safety response threshold, or when it is predicted that the maximum allowable slope constraint will be violated, the current stage constraint is immediately released and a local replanning from the current position to the subsequent skeleton node or target point is triggered, and the newly generated local path is smoothly integrated into the global trajectory.

[0040] Beneficial effects

[0041] 1. By constructing a terrain risk manifold field that includes multi-dimensional features such as slope and roughness, the method of this invention provides a continuous quantitative evaluation standard for environmental risks, guides the path to avoid areas with drastic terrain changes, and improves navigation safety.

[0042] 2. By adopting a two-layer dimensionality reduction search architecture, the method of the present invention constrains the computational range of graph search in a high-resolution environment to the dynamic corridor generated by the coarse-resolution skeleton, thereby reducing the number of node expansions and computational overhead while ensuring the quality of path planning.

[0043] 3. By using a gated replanning mechanism based on the incremental gradient of forward-looking risks, the method of this invention has the ability to dynamically respond to local unknown high-risk terrain during navigation, thus improving adaptability in online environments. Attached Figure Description

[0044] Figure 1 A flowchart of an autonomous underwater vehicle path planning method guided by terrain risk manifold and two-layer graph;

[0045] Figure 2 This is a flowchart illustrating the construction of a global terrain risk manifold field in an embodiment of the present invention;

[0046] Figure 3 This is a flowchart of the two-layer diagram planning method of coarse-layer skeleton-fine-layer corridor in an embodiment of the present invention;

[0047] Figure 4 This is a comprehensive terrain risk heat map and a schematic diagram of the path obstacle avoidance effects of the present invention and the comparison algorithm;

[0048] Figure 5 This is a schematic diagram of the spatial location of the three-dimensional underwater terrain and the optimal path generated by this invention. Detailed Implementation

[0049] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0050] In this embodiment, the offline simulation hardware and software testing platform uses a real open-source underwater terrain dataset as the test benchmark. This dataset originates from a real marine geological survey conducted between 2018 and 2023 using an autonomous underwater vehicle (AUV) equipped with a multibeam sonar system in the Carmel Canyon area of ​​Monterey Bay, California (approximately 36.5°N, 121.9°W). This test area features dramatic depth variations, dispersed rock distribution, and complex geological features including volcanic seamounts, sedimentary deposits, and submarine canyon walls. Stored in a three-dimensional mesh matrix, it is specifically used to evaluate and benchmark the obstacle avoidance performance of autonomous navigation algorithms in complex environments. This embodiment sets the following AUV mission dynamics constraints: maximum allowable slope threshold... Safe slope threshold The absolute hard upper limit of risk is set at 0.95.

[0051] like Figure 1 As shown, the autonomous underwater vehicle path planning method based on terrain risk manifold and two-layer graph guidance of the present invention includes:

[0052] Step S1: Obtain underwater topographic elevation grid data containing three-dimensional coordinates, and adaptively estimate the physical step size by calculating the median of the spatial difference between adjacent grids to complete the standardization of the data spatial scale;

[0053] Step S2: Extract terrain slope, roughness, depth deviation and local connectivity features, and construct a global continuous terrain risk manifold field through robust normalization method and weighted fusion algorithm;

[0054] Step S3: Downsample the terrain risk manifold to construct a coarse semantic graph, and perform global coarse-resolution skeleton path planning on the coarse semantic graph based on a composite cost function that includes risk and constraint penalties;

[0055] Step S4: Densify the coarse-resolution skeleton path and map it back to the high-resolution mesh, generate a fine-layer corridor mask with dynamic boundaries along the path, and perform a constraint refinement search within the space defined by the corridor mask.

[0056] Step S5: Apply sliding filter to smooth the independent dimensions of the discrete path nodes obtained by the search, and output a finely planned path after full-cycle dynamic constraint verification, so as to use the autonomous navigation control of the underwater vehicle.

[0057] Step S6: During the underwater vehicle's navigation along the path, continuously monitor the incremental gradient of risks along the path within the forward-looking window. When the safety response threshold is exceeded, perform local backoff or replanning to ensure dynamic adaptive safety.

[0058] Example 1

[0059] Step 1: Obtain the input 3D underwater terrain elevation grid data. In this embodiment, the input matrix grid resolution is 1000×1000 points, the spatial physical coverage is approximately 1 km × 1 km (1,000,000 square meters), the theoretical grid spacing is 1 m × 1 m, and the vertical accuracy is ±0.5 m. The actual grid elevations (depth Z-values) are distributed between 1.4794 m and 76.8312 m, with a mean of 13.8770 m. The physical step size is adaptively estimated by calculating the median of the spatial differences between adjacent grids in the corresponding dimensions of the X and Y matrices. The calculated actual grid step size in this embodiment is... , (Unit: meters), thus completing the standardization of data spatial scale.

[0060] Step 2: As Figure 2 As shown, on the standardized base map, the slope matrix, Laplacian roughness matrix, and depth-to-median deviation matrix are calculated sequentially. A connectivity penalty matrix is ​​calculated by generating a binary passable mask and applying a convolution kernel of a preset dimension (e.g., 3×3) for local smoothing. The four feature matrices are then robustly normalized using their 1% to 99th percentiles to filter out outliers. Finally, preset weights (slope weights are set in this embodiment) are used... Roughness weight Depth deviation weight Connectivity weights Construct a static, global terrain risk manifold field. The comprehensive terrain risk heatmap and path avoidance performance are shown below. Figure 4 As shown, the area of ​​dramatic topographic change and steep slope canyons ( Figure 4 High-risk areas (those with darker colors) were accurately quantified and identified.

[0061] Step 3: As Figure 3 As shown, on the 1000×1000 high-resolution global terrain risk manifold field obtained in step 2, a downsampling factor is introduced for block averaging to construct a 200×200 low-resolution coarse semantic map. Within the feasible region defined by the coarse map, an A* heuristic search is performed using a composite cost function, combining the mixed risk and the quadratic penalty of terrain slope truncation, to quickly solve for a low-resolution global coarse skeleton path that avoids macroscopic high-risk areas.

[0062] Step 4: As Figure 3As shown, the low-resolution global coarse skeleton path from the previous step is mapped back to a 1000×1000 high-resolution grid, and linear densification interpolation is performed between nodes. The radius of the interpolated skeleton trajectory is expanded outward by 24 grid cells to generate a dynamic corridor mask. Within the subdivision space defined by this mask, a constrained refinement search is performed again using the composite cost function described in step 3 to obtain a fine-layer feasible path that fits the high-resolution terrain safely. If the search fails within this corridor, the mask is automatically discarded, and a cost-constrained graph search is performed again in the entire feasible region.

[0063] Step 5: Extract the X, Y, and Z coordinates of the path points in the fine-grained layer. Apply a moving average filter with a window length of 7 to perform one-dimensional smoothing, while simultaneously employing an edge value replication strategy to eliminate jagged lines caused by the grid search. Perform dynamic constraint verification and calculation on the smoothed path, including slope exposure analysis and yaw angle cost assessment. The final output is as follows: Figure 5 The diagram shows a three-dimensional planning trajectory. Figure 5 The dark solid line represents the ground-hugging safe path generated by the TRM-Flow algorithm of this invention.

[0064] Step 6: During the online deployment phase, the vehicle control system will continuously calculate the risk increment gradient within the forward local window along this planned path. When a sharp increase in terrain risk is detected, a backtracking mechanism will be triggered to perform local replanning.

[0065] To objectively verify the effectiveness and performance advantages of this invention, this embodiment further sets up five batch experiments with different spatial spans at the start and end points. The specific local Cartesian coordinates (unit: meters) settings for the test scenarios are as follows:

[0066] Case-0: Start point (20.0, 20.0), End point (980.0, 980.0)

[0067] Case-1: Starting point (20.0, 980.0), Ending point (980.0, 20.0)

[0068] Case-2: Starting point (100.0, 100.0), Ending point (900.0, 850.0)

[0069] Case-3: Starting point (150.0, 850.0), Ending point (850.0, 150.0)

[0070] Case-4: Starting point (80.0, 500.0), Ending point (920.0, 500.0)

[0071] Three existing planning algorithms are introduced as baseline control groups: Baseline-A (single-layer shortest path A*), Baseline-B (single-layer risk A*), and Baseline-C (arbitrary angle single-layer Theta*).

[0072] To ensure the objectivity and rigor of the evaluation, the statistical evaluation indicators in this embodiment are divided into two categories: general physical standards and quantitative standards specific to this invention.

[0073] Among them, path length (length_m), maximum slope exposure (slope_max_deg), and planning time (runtime_sec) are common geometric and computational evaluation indicators in the field of robot path planning; while average risk (risk_mean), maximum risk (risk_max), and safe segment ratio (safe_ratio) are environmental safety evaluation indicators derived from the terrain risk manifold site constructed in step 2.

[0074] To evaluate the overall effectiveness of the aforementioned indicators, this embodiment introduces a comprehensive scoring mechanism. The comprehensive score has a maximum of 100 points and is primarily calculated by weighting and penalizing the redundancy ratio of the path length, average environmental risk, and the cost of path yaw angles. The specific calculation logic is as follows: path length penalty is calculated based on the theoretical straight-line Euclidean distance; risk penalty is calculated based on the average risk manifold value at path points; and yaw angle penalty is calculated based on the cumulative value of the difference in heading angles between adjacent path points. Final score. .

[0075] The statistical results (mean ± standard deviation) of the offline simulation of the system under five scenarios are shown in Table 1:

[0076]

[0077] Table 1

[0078] Table 1 shows that compared to the single-layer shortest path A* algorithm, this embodiment improves the average risk by 68.51%, and the safe segment ratio increases from 60.78% to 97.78%. Compared to the single-layer risk A* algorithm, this embodiment reduces the planning time from an average of 7.087s to 1.271s while maintaining the same safety evaluation. Compared to the arbitrary angle Theta* algorithm, although this embodiment is slightly lower in a few smoothing indices, it effectively avoids its high computational cost of 12.551s, proving that the two-layer graph planning architecture has extremely excellent engineering implementation value and computational cost economy.

[0079] This invention effectively solves the technical challenge of balancing path planning safety and real-time computation in complex seabed environments by constructing a risk manifold field that integrates multi-dimensional terrain features and introducing a two-layer dimensionality reduction search architecture. The risk manifold field provides global, continuous hazard avoidance guidance for the vehicle, enabling it to proactively avoid high-risk areas such as steep slopes and extremely rugged terrain. The two-layer graph guidance mechanism precisely limits the high-resolution graph search overhead in large-scale environments to local safe corridors, greatly improving the algorithm's computational efficiency and engineering deployment feasibility while strictly ensuring the dynamic safety baseline of underwater vehicles.

[0080] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A path planning method for autonomous underwater vehicles based on terrain risk manifold and two-layer graph guidance, characterized in that, Includes the following steps: Step 1: Obtain the elevation grid data of the underwater topography, construct a basic environmental map matrix including X coordinates, Y coordinates and corresponding elevation Z, and adaptively estimate the physical step size of the grid by calculating the median absolute value of the spatial difference between adjacent grids in the corresponding dimensions of the X matrix and Y matrix respectively. Step 2: Extract multi-dimensional terrain features based on the basic environment map matrix, and perform robust normalization on the multi-dimensional terrain features to construct a global terrain risk manifold field; Step 3: Downsample the global terrain risk manifold field and construct a low-resolution coarse semantic map, and plan a global coarse skeleton path within the feasible region of the coarse semantic map; Step 4: Map the global coarse skeleton path back and generate a dynamic corridor mask, and perform a refined path search within the feasible region defined by the dynamic corridor mask; Step 5: Extract the corresponding three-dimensional coordinate sequence from the refined path points obtained by the search and use a sliding average filter with a preset window size to perform independent one-dimensional smoothing on each dimension of X, Y, and Z. Eliminate the jagged edges of the grid search by using an edge value copying filling strategy. Then perform dynamic and safety constraint verification along the smoothed path points. Step 6: During the autonomous underwater vehicle's navigation along the planned path, gating replanning monitoring is implemented. By rolling the calculation of the incremental gradient of risks along the path within the forward-looking local window, the danger of unknown or abrupt terrain ahead is predicted in real time, and finally a planned path with dynamic adaptive capabilities is output.

2. The path planning method for autonomous underwater vehicles based on terrain risk manifold and two-layer graph guidance as described in claim 1, characterized in that: When the median absolute value is not a finite value or is not a positive number, a preset default step size constant is set.

3. The autonomous underwater vehicle path planning method based on terrain risk manifold and two-layer graph guidance as described in claim 1, characterized in that: Step 2 specifically includes the following steps: Step 2.1: Calculate the spatial gradient of elevation Z in the x and y directions to obtain the local slope angle, as shown in the following formula: ， wherein, is a local slope angle, and the slope feature is obtained by robustly normalizing the local slope angle. Step 2.2: Calculate the absolute value of the Laplace operator based on the second derivative of elevation Z, using the following formula: ， wherein, is the absolute value of the Laplacian operator, and is the local roughness proxy indicator, and is robustly normalized to obtain the roughness feature; Step 2.3: Calculate the absolute deviation of the local depth relative to the global median, and perform robust normalization to obtain the depth deviation feature; Step 2.4: Generate a binary passable mask based on whether the local slope angle is less than the maximum allowable slope. Use a convolution kernel of a preset dimension to locally smooth the mask to calculate the local continuous passability ratio, and take its complementary value as the connectivity penalty feature. Step 2.5, linearly weight and fuse the slope feature, roughness feature, depth deviation feature and connectivity penalty feature according to the set weight, and truncate to closed interval, construct the global terrain risk manifold field, and the formula is as follows: ， where S is a slope feature, R is a roughness feature, E is a depth deviation feature, C is a connectivity penalty feature, is a global terrain risk manifold field, , , , are set weights for the respective features.

4. The terrain-based risk manifold and double-layer graph-guided autonomous underwater vehicle path planning method of claim 2 or 3, wherein: The robust normalization process specifically involves extracting the 1% to 99% quantiles of terrain feature data in each dimension as upper and lower bounds for linear mapping, thereby eliminating the influence of extreme outliers.

5. The terrain-based risk manifold and double-layer graph-guided autonomous underwater vehicle path planning method of claim 3, wherein: Step 3 specifically includes the following steps: Step 3.1: Introduce a downsampling factor and use a block averaging strategy to downsample the global terrain risk manifold field, local slope angle, and local ruggedness proxy index to construct a low-resolution coarse-layer map. Step 3.2: Within the feasible region of the low-resolution coarse-layer graph, a heuristic graph search algorithm is used to plan a global coarse skeleton path, where the node transition cost function is a composite cost function, designed as follows: ， where 1.0 is the base step cost, are the global terrain risk manifold field, local slope angle and normalized local ruggedness proxy indicator after down-sampling, respectively, is the cost penalty hyper-parameter.

6. The path planning method for autonomous underwater vehicles based on terrain risk manifold and two-layer graph guidance as described in claim 5, characterized in that: Step 4 specifically includes the following steps: Step 4.1: Map the global coarse skeleton path back to the original high-resolution fine-layer map, and perform densification processing between adjacent skeleton nodes using a linear interpolation algorithm. Step 4.2: Extend the preset pixel radius along the densified skeleton nodes to generate a dynamic corridor mask. Step 4.3: Within the intersection of the dynamic corridor mask and the original high-resolution feasible region, perform a refined graph search using the same composite cost function as the coarse-layer search to obtain the refined-layer feasible path. Step 4.4: If no connected path can be found within the space defined by the dynamic corridor mask, a rollback mechanism is triggered, the corridor mask constraint is automatically discarded, and the constraint graph search is re-executed within the high-resolution feasible region of the entire graph.

7. The terrain-based risk manifold and double-layer graph-guided autonomous underwater vehicle path planning method of claim 1, wherein: The dynamics and safety constraint verification includes assessing whether the maximum slope exposure along the route exceeds the dynamics threshold, calculating the average and maximum risk limits of the path, and calculating the cumulative yaw cost of adjacent heading angles.

8. The path planning method for autonomous underwater vehicles based on terrain risk manifold and two-layer graph guidance as described in claim 1, characterized in that: When the risk increment gradient exceeds the set safety response threshold, or when it is predicted that the maximum allowable slope constraint will be violated, the current stage constraint will be released immediately and a local replanning from the current position to the subsequent skeleton node or target point will be triggered, and the newly generated local path will be smoothly integrated into the global trajectory.

Images