Recursive long-endurance running method for unmanned surface vehicle based on bimodal observation
By combining dual-modal observation and deep reinforcement learning, the problem of coordinating global path tracking and local collision avoidance in complex waters for unmanned surface vessels was solved, achieving efficient long-distance autonomous navigation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-04-07
AI Technical Summary
Existing path planning studies struggle to effectively address local uncertainties in complex waters during long voyages. Local collision avoidance decisions lack consideration for the globally optimal path, leading to frequent deviations of unmanned surface vessels from the global reference path, increasing voyage redundancy and energy consumption.
A recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation is adopted. The dual-modal observation space enhances the agent's perception of environmental details. By combining image modality and geometric modality observation, a local-global cooperative path planning strategy is constructed, and a deep reinforcement learning algorithm is used to optimize path decision.
It achieves coordinated optimization of global path tracking and local dynamic obstacle avoidance during long-distance navigation of unmanned surface vessels, solves the problem of slow algorithm convergence, and improves the efficiency of path planning and energy consumption management.
Smart Images
Figure CN121804473A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous navigation technology for unmanned surface vessels, specifically to a recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation. Background Technology
[0002] As a key vehicle for promoting the development of intelligent marine equipment, unmanned surface vessels (USVs) have become a research hotspot in the field of intelligent ships due to their long-distance autonomous operation capabilities. With the increasing complexity of operational tasks, how to achieve efficient coordination between global path tracking and local collision avoidance planning during long voyages places higher demands on USV navigation systems.
[0003] Existing path planning research largely relies on structured environment assumptions, employing simplified rectangular and circular models to model static obstacles. This neglects the irregular geometric properties of unstructured obstacles such as shorelines and reefs in real waters, leading to a lack of modeling accuracy and difficulty in effectively addressing local uncertainties in complex waters. Furthermore, existing navigation architectures often lack long-term consideration of the globally optimal path in local collision avoidance decisions, causing unmanned surface vessels (USVs) to frequently deviate from the global reference path when avoiding dynamic obstacles. This disconnect between local and global planning results in redundancy in the navigation path and increased energy consumption, becoming a technical bottleneck restricting USVs from achieving long-distance, efficient, and autonomous operations. Summary of the Invention
[0004] The purpose of this invention is to provide a recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation. By enhancing the agent's perception of environmental details through dual-modal observation, this method guides efficient autonomous navigation in differentiated long-distance operation scenarios.
[0005] This invention adopts the following technical solution: a recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation, comprising the following steps:
[0006] S1. Obtain the global electronic nautical chart of the area to be navigated, set the starting and ending coordinates of the unmanned surface vessel, and define the training batch;
[0007] S2. Generate a global reference path from the starting coordinates to the ending coordinates of the unmanned surface vessel based on the global path planning algorithm; decouple the global reference path using Manhattan distance constraints according to the distance length threshold to generate a series of recursive local sub-target points, thus obtaining a set of local sub-target points.
[0008] S3. Construct a local simulation training environment based on local sub-target points, define training segments, and randomly initialize the position and velocity vectors of dynamic obstacles for non-static obstacle blank areas in local space.
[0009] S4. Based on the configuration of shipborne sensors, generate simulation observation data, which includes the unmanned surface vessel's own position and velocity information, as well as the geometric distance measurement information between the unmanned surface vessel and obstacles.
[0010] S5. For the global reference path tracking task of long-endurance navigation, construct an image modal observation space and encode the input information of the edge features of static obstacles and the details of the reference path trajectory.
[0011] S6. For long-endurance navigation local collision avoidance planning tasks, construct a geometric modal observation space and encode the input information of the spatiotemporal motion state of dynamic obstacles;
[0012] S7. Define the reward function for regional regression reward and collision penalty, and use the value-distributed reinforcement learning algorithm to process the input information obtained in steps S5 and S6 to balance local collision avoidance and global tracking tasks.
[0013] S8. Based on the training segments, recursively initialize the sub-regions, and use the experience playback strategy to train the deep reinforcement learning model until convergence, thereby obtaining the local-global cooperative path planning strategy for the unmanned surface vessel.
[0014] S9. Verify the collaborative performance of the local-global collaborative path planning strategy for unmanned surface vessels in batches to achieve recursive long-term navigation of unmanned surface vessels.
[0015] Furthermore, in step S1, the global electronic nautical chart includes prior information on static obstacles such as reefs, shoals, and restricted areas within the navigation area.
[0016] Obtain static satellite maps of the actual aquatic environment and delineate the starting latitude and longitude coordinates of the global electronic nautical chart in the navigation coordinate system. and the latitude and longitude coordinates of the destination To move unmanned surface vessels from Drive to The process is defined as batch ;in, This indicates the latitude value of the coordinate. This indicates the longitude value of the coordinate.
[0017] Starting point latitude and longitude coordinates of the global electronic nautical chart Using the origin as the coordinate system, project the location of the static obstacle onto the carrier coordinate system and initialize the starting coordinates of the unmanned surface vessel. coordinates of the endpoint .
[0018] Furthermore, in step S2, Dijkstra's algorithm is used to generate the coordinates from the starting point of the unmanned surface vessel. Point to the destination coordinates The global reference path, which contains the set of coordinate points. , , Indicates the first in the global reference path Path coordinates, , ;
[0019] Calculate the Manhattan distance node by node, based on the distance length threshold. The global reference path is divided into continuous sub-segments, starting from the set of coordinate points. By selecting the start and end nodes of a sub-segment, a set of local sub-target points can be obtained. , , Represents the set of local sub-target points. Path coordinates.
[0020] Furthermore, in step S3, a random operator is determined from the local sub-target point set. The process of an unmanned surface vessel (USV) traveling from the starting point to the ending point of a local sub-region is defined as a segment, where two consecutive sub-target points are recursively selected as the starting and ending coordinates of the local sub-region. ;
[0021] For the fragment The non-static obstacle blank area is randomly initialized with the position and velocity information of dynamic obstacles. For the first... A dynamic obstacle, whose properties are defined as follows: ;in, and They represent the first A dynamic obstacle along shaft and Axis starting point coordinates, and They represent the first A dynamic obstacle along shaft and coordinates of the endpoint of the axis Indicates the first The velocity vector of a dynamic obstacle.
[0022] Furthermore, in step S4, based on the shipborne positioning system, obtain... Real-time location information of unmanned surface vessels Based on shipborne speed measurement devices, obtain Real-time surface unmanned surface vessel speed , unmanned surface vessel On-axis velocity components, unmanned surface vessel On-axis velocity components;
[0023] Based on a shipborne ranging system, and simulating lidar signals, geometric ranging information of an unmanned surface vessel relative to obstacles is obtained within a local sub-region. , Indicates the first Distance values measured by a single lidar beam.
[0024] Furthermore, in step S5, for the long-endurance global reference path tracking task, an image modal observation space is constructed as an input image; the set of coordinate points contained in the global reference path is... Overlaying it with the global static map yields a tracking sub-target guidance map;
[0025] by Real-time location information of unmanned surface vessels Centered on the target, capture the target tracking guidance map. Grid map of various sizes Extraction using convolutional neural networks The static features are used to guide the unmanned surface vessel to navigate along a global reference path, obtaining the edge features of static obstacles and the details of the reference path trajectory.
[0026] Furthermore, in step S6, for the long-endurance navigation local collision avoidance planning task, a geometric modal observation space is constructed as a vector input; using the backward extension of the unmanned surface vessel's heading angle as the initial scanning line, the geometric ranging information of the unmanned surface vessel relative to obstacles in a local sub-region is determined. Based on the inherent attributes of the unmanned surface vessel, self-information observation is determined. Independent multilayer perceptrons are used to process geometric ranging information separately. and self-information observation The dynamic characteristics guide the unmanned surface vessel to avoid obstacles by combining its own situation, and obtain the spatiotemporal motion state of the dynamic obstacles.
[0027] Furthermore, in step S7, the reward function is determined. This includes the reward value for the unmanned surface vessel reaching a local endpoint, the penalty value for colliding with obstacles, and the penalty value for the time step. ;
[0028] definition Time-to-Destination Distance Difference Reward Function Obstacle spacing reward function Global path regression reward function ;in, Indicates hyperparameters, Represents the regression distance value;
[0029] An independent network architecture is used to extract input information from the image modal observation space and the geometric modal observation space. After feature aggregation, the cumulative reward distribution information is obtained by value distributed reinforcement learning algorithm to guide the motion decision of unmanned surface vessels.
[0030] Furthermore, in step S8, for the training batch , Indicates the first A training segment, recursively initialize the training segment. Sub-region dynamic obstacle information; unmanned surface vessel in training segments Successfully reached the endpoint of the local sub-region Define identifier Training clips by Real-time location information of unmanned surface vessels Starting position For local endpoints;
[0031] Unmanned surface vessels in training footage Define an identifier when a collision or timeout occurs within the system. Training clips by Starting position This is a local endpoint.
[0032] Furthermore, in step S9, during the verification process, the unmanned surface vessel (USV) is used for training batches... There are identifiers in it. When this happens, subsequent sub-fragments will no longer be executed recursively, the current batch will be terminated, and this will be recorded. ;
[0033] When training batch All sub-fragments satisfy The unmanned surface vessel reached its destination. ,remember .
[0034] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:
[0035] 1. This invention utilizes a dual-modal state-space representation method that integrates image window observation and geometric perception observation. It designs a regression-based reward function to guide local path planning to closely approximate the global reference path, thereby achieving coordinated optimization of global path tracking and local dynamic obstacle avoidance during long-distance navigation of unmanned surface vessels.
[0036] 2. This invention fully considers the slow convergence problem of deep reinforcement learning models under long-distance navigation missions, and proposes a distance-constrained regionalized recursive architecture. By dynamically selecting key core nodes, the long-distance global path is intelligently divided into a variable number of navigation sub-regions, and a navigation process is executed recursively: sub-region path tracking, local real-time collision avoidance, and smooth transition between adjacent regions. This not only solves the algorithm scalability problem under paths of variable length, but also effectively alleviates the dilemma of local tracks deviating from the global reference path. Attached Figure Description
[0037] Figure 1 This is a flowchart illustrating the overall implementation of the present invention.
[0038] Figure 2 This is a schematic diagram of the image modal observation space of the present invention.
[0039] Figure 3 This is a schematic diagram of the geometric modal observation space of the present invention.
[0040] Figure 4 This is a diagram illustrating the effect of local-global collaborative path planning in an embodiment of the present invention. Detailed Implementation
[0041] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.
[0042] To achieve the above objectives, this invention proposes a recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation, such as... Figure 1 As shown, at the algorithm input level, this method constructs a dual-mode dual-branch data processing architecture: branch one receives data as follows... Figure 2 The image shown illustrates the edge features and reference path trajectory details of static obstacles generated in the modal observation space, as received by branch two. Figure 3 The spatiotemporal motion state of dynamic obstacles generated in the geometric modal observation space is shown, and features are extracted separately using independent neural networks. To address the technical challenge of local collision avoidance paths easily deviating from the global reference path, a recursive sub-region working mode is introduced, decoupling the long-distance global path into several continuous sub-segments, and using these as the basis for training a distributed reinforcement learning algorithm. This strategy effectively improves the algorithm's convergence speed while providing technical support for real-time and efficient local-global cooperative path planning for unmanned surface vessels in unstructured environments. The specific steps are as follows:
[0043] S1. Obtain the global electronic nautical chart of the area to be navigated, set the start and end coordinates of the unmanned surface vessel, and define the training batch; specifically:
[0044] The global electronic nautical chart includes prior information on static obstacles such as reefs, shoals, and restricted areas within the navigation area;
[0045] Obtain static satellite maps of the actual aquatic environment and delineate the starting latitude and longitude coordinates of the global electronic nautical chart in the navigation coordinate system. and the latitude and longitude coordinates of the destination To move unmanned surface vessels from Drive to The process is defined as batch ;in, This indicates the latitude value of the coordinate. This indicates the longitude value of the coordinate.
[0046] Starting point latitude and longitude coordinates of the global electronic nautical chart Using the origin as the coordinate system, project the location of the static obstacle onto the carrier coordinate system and initialize the starting coordinates of the unmanned surface vessel. coordinates of the endpoint .
[0047] S2. Generate a global reference path from the starting coordinates to the ending coordinates of the unmanned surface vessel based on a global path planning algorithm; decouple the global reference path using Manhattan distance constraints based on a distance length threshold to generate a series of recursive local sub-target points, thus obtaining a set of local sub-target points; specifically:
[0048] Using Dijkstra's algorithm to generate the starting coordinates of the unmanned surface vessel. Point to the destination coordinates The global reference path, which contains the set of coordinate points. , , Indicates the first in the global reference path Path coordinates, , ;
[0049] Calculate the Manhattan distance node by node, based on the distance length threshold. The global reference path is divided into continuous sub-segments, starting from the set of coordinate points. By selecting the start and end nodes of a sub-segment, a set of local sub-target points can be obtained. , , Represents the set of local sub-target points. Path coordinates.
[0050] S3. Construct a local simulation training environment based on local sub-target points, define training segments, and randomly initialize the position and velocity vectors of dynamic obstacles in the blank areas of non-static obstacles in the local space; specifically:
[0051] Determine the random operator from the local sub-target point set. The process of an unmanned surface vessel (USV) traveling from the starting point to the ending point of a local sub-region is defined as a segment, where two consecutive sub-target points are recursively selected as the starting and ending coordinates of the local sub-region. ;
[0052] For the fragment The non-static obstacle blank area is randomly initialized with the position and velocity information of dynamic obstacles. For the first... A dynamic obstacle, whose properties are defined as follows: ;in, and They represent the first A dynamic obstacle along shaft and Axis starting point coordinates, and They represent the first A dynamic obstacle along shaft and coordinates of the endpoint of the axis Indicates the first The velocity vector of a dynamic obstacle.
[0053] S4. Generate simulation observation data based on the configuration of shipborne sensors. The data includes the unmanned surface vessel's own position and velocity information, as well as the geometric distance measurement information between the unmanned surface vessel and obstacles; specifically:
[0054] Based on the shipborne positioning system, obtain Real-time location information of unmanned surface vessels Based on shipborne speed measurement devices, obtain Real-time surface unmanned surface vessel speed , unmanned surface vessel On-axis velocity components, unmanned surface vessel On-axis velocity components;
[0055] Based on a shipborne ranging system, and simulating lidar signals, geometric ranging information of an unmanned surface vessel relative to obstacles is obtained within a local sub-region. , Indicates the first Distance values measured by a single lidar beam.
[0056] S5. For the global reference path tracking task in long-endurance navigation, construct an image modal observation space and encode the input information of edge features of static obstacles and details of the reference path trajectory; specifically:
[0057] For long-endurance global reference path tracking tasks, an image modal observation space is constructed as input in image form; the set of coordinate points contained in the global reference path is... Overlaying it with the global static map yields a tracking sub-target guidance map;
[0058] by Real-time location information of unmanned surface vessels Centered on the target, capture the target tracking guidance map. Grid map of various sizes Extraction using convolutional neural networks The static features are used to guide the unmanned surface vessel to navigate along a global reference path, obtaining the edge features of static obstacles and the details of the reference path trajectory.
[0059] S6. For long-endurance navigation local collision avoidance planning tasks, construct a geometric modal observation space and encode the input information of the spatiotemporal motion state of dynamic obstacles; specifically:
[0060] For long-endurance navigation local collision avoidance planning tasks, a geometric modal observation space is constructed with vector input; using the backward extension of the unmanned surface vessel's heading angle as the initial scan line, the geometric ranging information of the unmanned surface vessel relative to obstacles in a local sub-region is determined. Based on the inherent attributes of the unmanned surface vessel, self-information observation is determined. Independent multilayer perceptrons are used to process geometric ranging information separately. and self-information observation The dynamic characteristics guide the unmanned surface vessel to avoid obstacles by combining its own situation, and obtain the spatiotemporal motion state of the dynamic obstacles.
[0061] S7. Define reward functions for region regression reward and collision penalty, and process the input information obtained in steps S5 and S6 using a value-distributed reinforcement learning algorithm to balance local collision avoidance and global tracking tasks; specifically:
[0062] Determine the reward function This includes the reward value for the unmanned surface vessel reaching a local endpoint, the penalty value for colliding with obstacles, and the penalty value for the time step. ;
[0063] definition Time-to-Destination Distance Difference Reward Function Obstacle spacing reward function Global path regression reward function ;in, Indicates hyperparameters, Represents the regression distance value;
[0064] An independent network architecture is used to extract input information from the image modal observation space and the geometric modal observation space. After feature aggregation, the cumulative reward distribution information is obtained by value distributed reinforcement learning algorithm to guide the motion decision of unmanned surface vessels.
[0065] S8. Based on training segments, sub-regions are recursively initialized, and a deep reinforcement learning model is trained until convergence using an experience replay strategy to obtain a local-global cooperative path planning strategy for the unmanned surface vessel; specifically:
[0066] For training batches , Indicates the first A training segment, recursively initialize the training segment. Sub-region dynamic obstacle information; unmanned surface vessel in training segments Successfully reached the endpoint of the local sub-region Define identifier Training clips by Real-time location information of unmanned surface vessels Starting position For local endpoints;
[0067] Unmanned surface vessels in training footage Define an identifier when a collision or timeout occurs within the system. Training clips by Starting position This is a local endpoint.
[0068] S9. Verify the collaborative performance of the local-global collaborative path planning strategy for unmanned surface vessels in batches to achieve recursive long-endurance navigation of unmanned surface vessels; specifically:
[0069] During the verification process, for training batches of unmanned surface vessels... There are identifiers in it. When this happens, subsequent sub-fragments will no longer be executed recursively, the current batch will be terminated, and this will be recorded. ;
[0070] When training batch All sub-fragments satisfy The unmanned surface vessel reached its destination. ,remember .
[0071] Example:
[0072] This invention relates to path planning technology for unmanned surface vessels (USVs), aiming to solve the problem of autonomous path planning for USVs in irregular real waters and with varying voyage distances. In this embodiment, combining the convergent parameters of a deep reinforcement learning network in a simulated environment, and using a real lake scene as the algorithm verification background, the specific implementation process and related parameter configurations for USVs performing cooperative tasks of local collision avoidance and global tracking are described in detail:
[0073] Step 1: Download a satellite map of a body of water, including information on static obstacles such as shorelines, islands, and docks, using Google Satellite Maps. (Using coordinates...) Using the origin as the reference point, the ship's self- From time s to The initial navigation trajectory within the time interval s is the global reference path.
[0074] Step 2: Overlay the static satellite map with the global reference path to construct a guidance map. Define the distance length threshold. The global reference path is broken down into a recursive set of local sub-target points. , .
[0075] Step 3: Initialize the dynamic obstacles in each sub-segment using the actual trajectories of two ships sailing on the water, and record their position and speed information in real time.
[0076] Step 4: Obtain the geometric ranging information of the reference vessel relative to the obstacle vessel based on the shipborne lidar, and obtain the position and velocity information of the reference vessel itself based on the shipborne inertial navigation unit and satellite navigation module.
[0077] Step 5: For the long-endurance global reference path tracking task, construct an image modal observation space, and encode the input information of the edge features of static obstacles and the details of the reference path trajectory. .
[0078] Step 6: Real-time ranging information based on the geometric ranging information of the reference ship relative to the obstacle in a local sub-region acquired by lidar. And referencing the ship's own motion state A geometric modal observation space was constructed to encode the spatiotemporal motion state of dynamic obstacles for local collision avoidance planning tasks during long-endurance navigation.
[0079] Step 7: Regress the reward function for the global reference path. hyperparameters Regression distance value Defined as the distance the reference vessel travels from leaving the global reference path to rejoining it, with the time penalty step size. When the distance between the reference vessel and the obstacle is less than Define distance scaling weights in meters. Calculate the obstacle spacing reward function .
[0080] Step 8: In algorithm performance testing, recursively execute the sub-fragments. The results showed that the reference vessel successfully avoided two dynamic obstacle vessels while only experiencing minor local path deviations, and ultimately arrived precisely at the preset endpoint of the global reference path, meeting the requirements. Plotted based on location information at each time point. Figure 4 The shown navigation trajectory Figure 4 (a) shows the effect of global reference path tracking. It can be seen that the solid line trajectory of the reference ship matches the navigation trend of the dashed line trajectory of the global reference path. Figure 4 (b) is a diagram showing the effect of local collision avoidance planning. It can be seen that the solid trajectory of the reference ship and the dashed trajectory of the two dynamic obstacle ships maintain a safe navigation distance in each sub-segment, indicating that the reference ship can achieve long-term navigation in a coordinated manner of global tracking and local collision avoidance, which effectively verifies the coordinated performance of the present invention at the level of local collision avoidance and global tracking.
[0081] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.
[0082] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation, characterized in that, include: S1. Obtain the global electronic nautical chart of the area to be navigated, set the starting and ending coordinates of the unmanned surface vessel, and define the training batch; S2. Generate a global reference path from the starting coordinates to the ending coordinates of the unmanned surface vessel based on the global path planning algorithm; decouple the global reference path using Manhattan distance constraints according to the distance length threshold to generate a series of recursive local sub-target points, thus obtaining a set of local sub-target points. S3. Construct a local simulation training environment based on local sub-target points, define training segments, and randomly initialize the position and velocity vectors of dynamic obstacles for non-static obstacle blank areas in local space. S4. Based on the configuration of shipborne sensors, generate simulation observation data, which includes the unmanned surface vessel's own position and velocity information, as well as the geometric distance measurement information between the unmanned surface vessel and obstacles. S5. For the global reference path tracking task of long-endurance navigation, construct an image modal observation space and encode the input information of the edge features of static obstacles and the details of the reference path trajectory. S6. For long-endurance navigation local collision avoidance planning tasks, construct a geometric modal observation space and encode the input information of the spatiotemporal motion state of dynamic obstacles; S7. Define the reward function for regional regression reward and collision penalty, and use the value-distributed reinforcement learning algorithm to process the input information obtained in steps S5 and S6 to balance local collision avoidance and global tracking tasks. S8. Based on the training segments, recursively initialize the sub-regions, and use the experience playback strategy to train the deep reinforcement learning model until convergence, thereby obtaining the local-global cooperative path planning strategy for the unmanned surface vessel. S9. Verify the collaborative performance of the local-global collaborative path planning strategy for unmanned surface vessels in batches to achieve recursive long-term navigation of unmanned surface vessels.
2. The recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation as described in claim 1, characterized in that, In step S1, the global electronic nautical chart includes prior information on static obstacles such as reefs, shoals, and restricted areas within the navigation area; Obtain static satellite maps of the actual aquatic environment and delineate the starting latitude and longitude coordinates of the global electronic nautical chart in the navigation coordinate system. Latitude and longitude coordinates of the destination To move unmanned surface vessels from Drive to The process is defined as batch ;in, This indicates the latitude value of the coordinates. This indicates the longitude value of the coordinate. Starting point latitude and longitude coordinates of the global electronic nautical chart Using the origin as the coordinate system, project the location of the static obstacle onto the carrier coordinate system and initialize the starting coordinates of the unmanned surface vessel. coordinates of the endpoint .
3. The recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation as described in claim 1, characterized in that, In step S2, Dijkstra's algorithm is used to generate the coordinates of the starting point of the unmanned surface vessel. Point to the destination coordinates The global reference path, which contains the set of coordinate points. , , Indicates the first in the global reference path Path coordinates points, , ; Calculate the Manhattan distance node by node, based on the distance length threshold. The global reference path is divided into continuous sub-segments based on the set of coordinate points. By selecting the start and end nodes of a sub-segment, a set of local sub-target points can be obtained. , , Represents the set of local sub-target points. Path coordinates.
4. The recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation according to claim 1, characterized in that, In step S3, a random operator is determined from the local sub-target point set. The process of an unmanned surface vessel (USV) traveling from the starting point to the ending point of a local sub-region is defined as a segment, where two consecutive sub-target points are recursively selected as the starting and ending coordinates of the local sub-region. ; For the fragment The non-static obstacle blank area is randomly initialized with the position and velocity information of dynamic obstacles. For the first... A dynamic obstacle, whose properties are defined as follows: ;in, and They represent the first A dynamic obstacle along shaft and Axis starting point coordinates, and They represent the first A dynamic obstacle along shaft and coordinates of the endpoint of the axis Indicates the first The velocity vector of a dynamic obstacle.
5. The recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation according to claim 1, characterized in that, In step S4, based on the shipborne positioning system, obtain... Real-time location information of unmanned surface vessels ; Based on shipborne speed measurement devices, obtain Real-time surface unmanned surface vessel speed , unmanned surface vessel On-axis velocity components, unmanned surface vessel On-axis velocity components; Based on a shipborne ranging system, and simulating lidar signals, geometric ranging information of an unmanned surface vessel relative to obstacles is obtained within a local sub-region. , Indicates the first Distance values measured by a single lidar beam.
6. The recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation according to claim 5, characterized in that, In step S5, for the long-endurance global reference path tracking task, an image modal observation space is constructed as an input image. The set of coordinate points contained in the global reference path Overlaying it with the global static map yields a tracking sub-target guidance map; by Real-time location information of unmanned surface vessels Centered on the target, capture the target tracking guidance map. Grid map of various sizes Extraction using convolutional neural networks The static features are used to guide the unmanned surface vessel to navigate along a global reference path, obtaining the edge features of static obstacles and the details of the reference path trajectory.
7. The recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation according to claim 5, characterized in that, In step S6, for the long-endurance navigation local collision avoidance planning task, a geometric modal observation space is constructed with vector input; using the backward extension of the unmanned surface vessel's heading angle as the initial scanning line, the geometric ranging information of the unmanned surface vessel relative to obstacles in a local sub-region is determined. Based on the inherent attributes of the unmanned surface vessel, self-information observation is determined. Independent multilayer perceptrons are used to process geometric ranging information separately. and self-information observation The dynamic characteristics guide the unmanned surface vessel to avoid obstacles by combining its own situation, and obtain the spatiotemporal motion state of the dynamic obstacles.
8. The recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation according to claim 1, characterized in that, In step S7, the reward function is determined. This includes the reward value for the unmanned surface vessel reaching a local endpoint, the penalty value for colliding with obstacles, and the penalty value for the time step. ; definition Time-to-Destination Distance Difference Reward Function Obstacle spacing reward function Global path regression reward function ;in, Indicates hyperparameters, Represents the regression distance value; An independent network architecture is used to extract input information from the image modal observation space and the geometric modal observation space. After feature aggregation, the cumulative reward distribution information is obtained by value distributed reinforcement learning algorithm to guide the motion decision of unmanned surface vessels.
9. The recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation according to claim 1, characterized in that, In step S8, for the training batch , Indicates the first A training segment, recursively initialize the training segment. Sub-region dynamic obstacle information; unmanned surface vessel in training segments Successfully reached the endpoint of the local sub-region Define identifier Training clips by Real-time location information of unmanned surface vessels Starting position For local endpoints; Unmanned surface vessels in training footage Define an identifier when a collision or timeout occurs within the system. Training clips by Starting position This is a local endpoint.
10. The recursive long-endurance navigation method for unmanned surface vessels based on dual-modal observation according to claim 1, characterized in that, In step S9, during the verification process, the unmanned surface vessel (USV) performs verification on the training batch. There are identifiers in it. When this happens, subsequent sub-fragments will no longer be executed recursively, the current batch will be terminated, and this will be recorded. ; When training batch All sub-fragments satisfy The unmanned surface vessel reached its destination. ,remember ; in, This indicates the latitude value of the coordinates. This indicates the longitude value of the coordinate. This represents the total number of local sub-target points.