CN-122018517-A - Space-time tensor task priority local planning method for unmanned surface vessel

CN122018517ACN 122018517 ACN122018517 ACN 122018517ACN-122018517-A

Abstract

The invention provides a space-time tensor task priority local planning method for an unmanned surface vessel, which belongs to the technical field of autonomous unmanned system navigation and control and comprises the steps of constructing a unified three-dimensional space-time tensor based on environment perception data, integrally coding static obstacle information and future prediction occupation information of a dynamic obstacle, supporting collision feasibility query of constant time complexity of any position and time combination, generating guidance reference information containing expected heading based on a time parameterized reference track, generating a candidate control action set based on a current motion state, performing collision constraint check on the candidate control action based on the three-dimensional space-time tensor, screening to obtain a safe action set meeting no collision constraint, selecting optimal control actions in the safe action set by taking deviation of minimized predicted ship heading and expected heading as a unique optimization target, and generating a local planning path based on the optimal control actions to control unmanned surface vessel navigation.

Inventors

YU WANNENG
WANG SHAOWEI
DAI TIANTIAN
LIAO WEIQIANG
LI SUWEN

Assignees

集美大学

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (10)

1. The space-time tensor task priority local planning method for the unmanned surface vessel comprises the steps of obtaining the current motion state of the unmanned surface vessel, a preset time parameterized reference track and environment perception data acquired by a ship-borne sensor, and is characterized by further comprising the following steps: The method comprises the steps of firstly, constructing a unified three-dimensional space-time tensor based on environment perception data, wherein the three-dimensional space-time tensor carries out integrated coding on static obstacle information and future prediction occupation information of dynamic obstacles, and supports collision feasibility inquiry of constant time complexity of any position and time combination; Generating guidance reference information containing expected heading based on the time parameterized reference track; Thirdly, generating a candidate control action set based on the current motion state, performing collision constraint verification on the candidate control actions based on the three-dimensional space-time tensor, screening to obtain a safety action set meeting no collision constraint, and selecting an optimal control action in the safety action set by taking the deviation of the minimum predicted ship course and the expected course as a unique optimization target; and step four, generating a local planning path based on the optimal control action, and controlling the unmanned surface vessel to navigate.
2. The method for space-time tensor task priority local planning of an unmanned surface vessel according to claim 1, wherein the three-dimensional space-time tensor is a multi-layer three-dimensional structure formed by expanding a space grid with the unmanned surface vessel as a center along a time dimension, and the preamble multi-layer stores dynamic obstacle occupation state information in the current moment and future prediction time domain, and the last layer stores static obstacle information.
3. The space-time tensor task priority local planning method for the unmanned surface vessel, according to claim 2, is characterized in that when a unified three-dimensional space-time tensor is constructed, morphological expansion operation is carried out twice on the initial occupied tensor, a physical safety boundary and a robust safety boundary are respectively generated, and finally three-level semantic assignment is carried out on the three-dimensional space-time tensor, wherein a value 0 corresponds to a free area, a value 1 corresponds to the physical safety boundary, and a value 2 corresponds to the robust safety boundary.
4. The method for space-time tensor task priority local planning of the unmanned surface vessel according to claim 1, wherein the candidate control actions are generated by taking the yaw angular velocity of the unmanned surface vessel at the current moment as a reference, selecting a reference action closest to the current yaw angular velocity in a preset discrete yaw angular velocity action space, and forming a candidate control action set together by a previous adjacent action and a next adjacent action of the reference action.
5. A space-time tensor task priority local planning method for an unmanned surface vessel according to claim 3, wherein the collision constraint verification rule is to determine that the candidate control action satisfies a collision-free constraint only when the space-time tensor value of a layer storing dynamic obstacle occupancy state information and a layer storing static obstacle information at the corresponding predicted time is 0, respectively, at the predicted vessel position corresponding to the candidate control action.
6. A space-time tensor task-first local planning method for an unmanned surface vessel as claimed in claim 1, wherein: the evaluation function corresponding to the optimization target is as follows: Wherein, the Is candidate action Is used for the evaluation value of (a), Predicted ship heading angles corresponding to the candidate actions, And selecting the candidate action with the minimum evaluation value as the optimal control action.
7. The space-time tensor task priority local planning method for the unmanned surface vessel, which is characterized by comprising the step of embedding a single-step decision process in a depth-priority multi-step search framework with backtracking for execution, wherein a state tree is gradually expanded in a preset planning time domain to generate a local planning path, and when a certain branch does not have candidate actions meeting collision-free constraint, backtracking is performed on the other candidate actions in the previous state until a complete feasible local planning path is generated or backtracking is performed to an initial state.
8. The space-time tensor task priority local planning method for the unmanned surface vessel according to claim 1, wherein when forward prediction is performed based on the candidate control actions, longitudinal speed adjustment is synchronously performed, namely, acceleration compensation is limited by maximum acceleration of the vessel when the expected speed in the guidance reference information is greater than the current speed, and deceleration adjustment is limited by maximum deceleration of the vessel when the expected speed is less than or equal to the current speed, so as to compensate time delay generated in the collision avoidance process.
9. A space-time tensor task-first local planning system for an unmanned surface vessel, comprising: The data acquisition module is used for acquiring the current motion state of the unmanned surface vessel, a preset time parameterized reference track and environment perception data acquired by the on-board sensor; The system comprises a space-time tensor construction module, a space-time tensor analysis module and a space-time tensor analysis module, wherein the space-time tensor construction module is used for constructing a unified three-dimensional space-time tensor based on environment perception data, and the three-dimensional space-time tensor carries out integrated coding on static obstacle information and future prediction occupation information of a dynamic obstacle and supports collision feasibility inquiry of constant time complexity of any position and time combination; the guidance reference generation module is used for generating guidance reference information containing a desired course based on the time parameterized reference track; a candidate action generating module for generating a candidate control action set based on the current motion state; the task priority decision module is used for carrying out collision constraint verification on the candidate control actions based on the three-dimensional space-time tensor, screening to obtain a safety action set meeting no collision constraint, taking the deviation between the minimum predicted ship course and the expected course in the safety action set as a unique optimization target, and selecting the optimal control action; and the navigation control module is used for generating a local planning path based on the optimal control action and controlling unmanned surface vessel navigation.
10. The space-time tensor task-first local planning system for the unmanned surface vessel according to claim 9, wherein the space-time tensor construction module is specifically configured to generate a two-dimensional occupation grid centered on the unmanned surface vessel by using laser radar point cloud projection, predict motion of dynamic obstacles by using a constant-speed model, and perform morphological dilation operation twice on the initial occupation tensor to generate a three-dimensional space-time tensor with three-level safety semantics.

Description

Space-time tensor task priority local planning method for unmanned surface vessel Technical Field The invention belongs to the technical field of autonomous unmanned system navigation and control, and particularly relates to a space-time tensor task priority local planning method for an unmanned surface vessel. Background With the continuous increase of the sea traffic density, the problems of navigation safety and operation efficiency are increasingly highlighted, and the rapid development of an autonomous navigation system is promoted. However, the local marine environment is still highly complex and uncertain. The coexistence of static and dynamic obstacles, environmental disturbance further increases the difficulty of autonomous navigation, and these factors often cause unmanned surface vessels to deviate from a given course during operation. In a collaborative or time constrained mission scenario, such as formation voyage, multi-vessel collaborative sampling, and arrival mission with time window constraints, the reference motion of the vessel is not only spatially defined, but also parameterized in the time dimension. Therefore, collision prevention is required to ensure the safety of the geometric layer, and the consistency of the task progress of the ship along the time parameterized track is also required to be maintained. If the vessel deviates from the reference trajectory during collision avoidance, and returns to the geometric path only after that, time tracking errors are accumulated, which may lead to a failure of the coordination, a delay in mission progress or a decline in mission efficiency. Therefore, the collision avoidance process needs to maintain consistency with the reference trajectory in both spatial and temporal dimensions. The existing partial collision prevention method of the unmanned surface vessel is generally based on geometric or kinematic path planning. Conventional methods such as a velocity barrier (VO) method, a dynamic window method (DWA), and an Artificial Potential Field (APF) -based planner are widely used in practice. However, these methods often have typical limitations such as being prone to local minima, inadequate adaptability in complex interaction scenarios, and poor track smoothness. In order to alleviate the above problems, various improvements have been proposed in recent years. For example, the potential field model is enhanced by introducing an environmental density and spatial risk measure to suppress ringing and avoid undesirable local convergence. And a learning mechanism is introduced into the DWA framework, so that the online self-adaptive adjustment of the evaluation weight is realized, and the manual parameter adjustment is reduced. In addition, the sampling planning method also improves the searching strategy and the track generation mechanism to accelerate the convergence speed and generate a smoother path. Although the method improves planning performance to a certain extent, the essence of the method still belongs to a planning driving paradigm, and the collision avoidance process is mainly modeled on a geometric or kinematic level. Meanwhile, deep Reinforcement Learning (DRL) -based planning methods are increasingly emerging, which reduce human intervention and promote policy flexibility through self-learning capabilities. However, such methods typically require a large amount of training data, and their generalization ability and robustness remain limited in the face of unknown dynamic environments or distribution variations. For this reason, some hybrid approaches combining DRL with the conventional model have also emerged in recent years. For example, reinforcement learning is combined with reciprocal velocity barrier (RVO) theory to more effectively address static and dynamic barriers. However, in general, whether conventional or learning-based, collision avoidance is mostly considered as an independent planning problem, without explicitly considering closed-loop trajectory tracking and task-level time constraints. Therefore, even if the ship can successfully detour around the obstacle and return to the reference path, the time deviation generated during collision avoidance may not meet the requirements of the time-sensitive task. Unlike the planning-based approach, another class of studies incorporate collision avoidance directly into the control layer. For example, there are schemes of VO and vector field methods as guidance laws to enable a ship to realize static obstacle avoidance while following a path, and schemes of improving line-of-sight (LOS) algorithms based on geometric relationships to realize obstacle avoidance. In such guidance law-based methods, collision avoidance is mainly achieved by guiding the vessel to return to the reference path, but the time progress is not explicitly constrained. In recent years, control methods based on optimization (such as nonlinear model predictive control, NMPC) have been used for unified tra