A path planning method for collaborative operation of multiple brush boat robots

Abstract

The present invention discloses a path planning method for the collaborative operation of multiple boat-cleaning robots, belonging to the field of robot path planning technology. The method comprises the following steps: initializing parameters and collecting real-time data collected by sensors mounted on the boat-cleaning robots; formulating a robot group priority strategy; rasterizing the hull surface using a grid contour method; generating an initial path for the boat-cleaning robot group using an improved #imgabs0# algorithm; adding a binary tree constraint tree for path planning; and when there are no obstacles on the line connecting the front and rear nodes of a certain node, deleting redundant nodes on the path, retaining only the starting and ending points and inflection points, then deleting redundant inflection points, extracting key nodes as intermediate target points, and outputting an optimal path solution. The technical solution of the present invention can improve the collaborative work efficiency of multiple boat-cleaning robots and solve the technical problems of traditional path planning in complex curved surface environments, such as the lag in dynamic obstacle avoidance response and the low efficiency of multi-machine collaboration.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot path planning, and in particular relates to a method for planning paths for collaborative operations of multiple brush boat robots. Background Art

[0002] With global shipping activity increasing, the maintenance of ship navigation equipment has become a critical issue in ensuring maritime efficiency. When ships of all types are immersed in seawater for extended periods, complex marine biota gradually develop in the submerged areas of the hull, forming a destructive biofouling layer. These biofouling organisms, such as barnacles and calcareous worms, secrete bioadhesives that form a strong bond with the hull's protective coating, making them difficult to remove with conventional cleaning methods.

[0003] The continued accumulation of this stubborn biofouling raises multiple technical challenges: First, changes in the hydrodynamic properties of the hull surface significantly increase navigation resistance, leading to a surge in fuel consumption. Second, the weight of the biomass not only alters the ship's load distribution but can also cause safety hazards such as center of gravity shift, severely impairing navigation stability. Finally, the electrochemical interaction between the deposits and the metal hull accelerates corrosion. Therefore, the development of an efficient hull bio-cleaning system has become a key development direction in modern ship maintenance technology, and technological breakthroughs have dual practical significance for improving shipping economy and navigation safety.

[0004] Ship-cleaning robots are a type of specialized robot that has emerged in the shipping and marine engineering sectors in recent years. Their development is closely tied to ship maintenance needs, evolving environmental regulations, and the advancement of automation technology. Ships navigate the ocean for extended periods, and marine organisms such as algae and shellfish are susceptible to attaching to their hulls (a process known as biofouling). This not only increases resistance, fuel consumption, and carbon emissions, but can also accelerate hull corrosion, threatening navigation safety. Traditional manual cleaning methods are inefficient, costly, and rely on underwater divers, posing safety risks. With the International Maritime Organization (IMO) tightening its energy efficiency and environmental regulations for ships (such as the Anti-fouling Systems Convention), the industry urgently needs more efficient and sustainable hull cleaning solutions, which has become the direct driving force behind the development of ship-cleaning robots.

[0005] In the field of underwater maintenance of ship-brushing robots, the traditional single-machine operation mode is limited by equipment motion space constraints, operation efficiency bottlenecks and trajectory conflicts. It is urgent to realize the innovation of the operation mode through multi-machine collaborative path planning technology.

[0006] Therefore, in order to address the key technical challenges commonly encountered in ship hull biofouling removal projects, such as cleaning blind spots, rising operation and maintenance costs, and insufficient adaptive capabilities for complex surfaces, research on collaborative operation path planning for multiple brushing robots has important engineering value. By constructing a dynamic task allocation model and a distributed conflict resolution mechanism, this technology can achieve spatiotemporal collaborative optimization of multi-machine operation trajectories, effectively eliminating coverage blind spots in traditional serial operation modes, and significantly reducing the risk of robotic arm interference through kinematic coupling planning. Compared with discrete operation schemes, collaborative path planning can systematically shorten the docking operation cycle and optimize the energy consumption distribution of the high-pressure water jet system, providing key technical support for the transformation of the ship's green maintenance system to an intelligent one. Summary of the Invention

[0007] In response to the above-mentioned problems existing in the prior art, the present invention proposes a path planning method for collaborative operation of multiple brush boat robots, which has a reasonable design, solves the shortcomings of the prior art, and has good effects.

[0008] A method for collaborative operation path planning of multiple brush boat robots includes the following steps:

[0009] Step 1: Initialize the parameters, including: hull surface map information, terrain obstacles and constraints, and the starting and ending point information of the ship washing robot;

[0010] Step 2: Collect real-time data from sensors on the ship-washing robot, including encoders, IMUs, and temperature and humidity sensors. The encoders collect the real-time speed of the ship-washing robot while it is working, the IMUs collect the position coordinates and posture information of the ship-washing robot, and the temperature and humidity sensors monitor changes in the temperature and humidity of the seawater.

[0011] Step 3: Determine the number of ship-washing robots, allocate the same type of robots according to the task size, and formulate a robot group priority strategy to ensure that there are no conflicts in the paths when multiple robots collaborate;

[0012] Step 4: Use the grid contour method to perform grid processing on the hull surface;

[0013] Step 5: Use the improved The algorithm generates the initial path for the ship-washing robot group;

[0014] Step 6: Add a binary tree constraint tree to detect whether there are conflicts among the solutions of each node in the initial path of the boat-cleaning robot population. If a conflict is detected in the current path, add constraints to each of the two child nodes generated for each conflict, and replan the affected path for the robot with low priority. The algorithm terminates when all paths are conflict-free and the total cost is optimal.

[0015] Step 7: Smoothly optimize the path and traverse all nodes on the robot path. When there are no obstacles on the line connecting the front and rear nodes of a certain node, delete the redundant nodes of the path and only keep the starting and ending points and inflection points. Then delete the redundant inflection points and extract the key nodes as the intermediate target points to output the optimal path solution.

[0016] Furthermore, the step 5 includes the following sub-steps:

[0017] Step 5.1, initialize the generated sets Open List and Close List. The Open List stores all nodes to be expanded, and the Close List stores all expanded nodes. Insert the start node into the Open List.

[0018] Step 5.2: When the Open List is not empty, select the node with the lowest total cost from the Open List. The total cost of the nodes The calculation formula is:

[0019] ;

[0020] in, is the actual path cost from the starting point to the current point, is the estimated path cost from the current point to the target point; Q-Learning algorithm is used to optimize , Using the Euclidean distance formula, The expression is:

[0021] ;

[0022] in, is the coordinate of the current node; is the coordinate of the starting node;

[0023] The node with the lowest total cost is removed from the Open List, added to the Close List, and used as the current node;

[0024] Step 5.3: If the current node is the target node, backtrack to the parent node to generate a path and end the algorithm. If the current node is not the target node, proceed to the next step.

[0025] Step 5.4. Find all the adjacent nodes of the current node. For each adjacent node, first determine whether it is in the Close List. If it is in the Close List, skip the node. If not, calculate the total cost of the adjacent nodes. ; Then determine whether it is in the Open List. If not, add it to the Open List and update the total cost of the adjacent nodes and the parent node. If it is, compare the current total cost and the total cost calculated previously , if the current total cost is smaller, then update the parent node and update the total cost of the adjacent nodes to ;

[0026] Step 5.5: Return to step 5.2 and continue the loop until the target node is found.

[0027] Furthermore, in step 5.2, the Q-Learning algorithm is used to optimize , including the following sub-steps:

[0028] Step 5.2.1: Initialize the Q-Learning algorithm parameters, obstacle information in the environment, number of robots, and coordinates of the starting and target points;

[0029] Construct the basic environment model for robot path planning;

[0030] Discretize the working area into uniform grids, and mark each grid as free space, obstacle or task target point;

[0031] The robot state space is defined as the two-dimensional coordinate (x, y) and the distribution of obstacles in its surrounding 3×3 neighborhood;

[0032] Define the robot's action space , The robot moves from the current grid unit to the adjacent Each action moves one step in each direction, corresponding to a set of eight neighborhood movement directions. Before each action is executed, feasibility verification must be performed: the target grid must be within the map boundary and not occupied by static obstacles;

[0033] Step 5.2.2: Use the R-Learning algorithm to obtain the optimal strategy by maximizing the reward value and introducing a collision avoidance reward and penalty mechanism. Specifically:

[0034] The total reward obtained by the boat-cleaning robot in the current state s after executing action a after selecting the strategy is composed of the collision avoidance reward and the step reward, which can be expressed as:

[0035] ;

[0036] in, is a sparse reward function, expressed as:

[0037] ;

[0038] is the distance reward function, expressed as:

[0039] ;

[0040] Where: is the reward coefficient, is the Euclidean distance between the starting point and the end point of the target;

[0041] During the training process, the Q value update follows the Bellman optimal equation, which is expressed as:

[0042] ;

[0043] Where, 、 is the state at the tth iteration and the t+1th iteration; 、 are the actions at the tth iteration and t+1th iteration respectively, and are learning rate and discount factor, respectively, with a value range of [0,1], For particles in state Next action The instant rewards you receive, The particle is in the state Take action The expected Q value of is the maximum expected future reward for all possible actions in the new state;

[0044] Step 5.2.3: Get the Q value of each state-action pair through Q-Learning training, traverse all grid nodes, and record the The maximum Q value of the grid nodes , store the maximum Q value result into the matrix, the matrix index is the node coordinate, the value is the maximum Q value corresponding to the node; embed the trained Q value into Algorithm, reconstruction heuristic function , the formula is:

[0045] ;

[0046] Where: Starting from a node, the maximum Q value that can be obtained by selecting the optimal action is the Q value, which represents the long-term benefits in historical experience; is the Euclidean distance from the current node to the target node, which is used to dynamically adjust the weight according to the distance from the node to the target; parameter is the attenuation coefficient, and its value range is [0.3,1].

[0047] Furthermore, step 6 includes the following sub-steps:

[0048] Step 6.1: Create a binary constraint tree based on the conflict search mechanism. Each constraint tree node N contains three pieces of information: the constraint set, the solution to the problem, and the cost of solving the problem at that node. The constraint set contains constraints for all the ship-farming robots in the problem.

[0049] Step 6.2: Algorithm initialization phase is improved The algorithm generates an unconstrained optimal path for each boat-cleaning robot to form an initial solution set;

[0050] Step 6.3: Start the conflict detection mechanism, using the 4-tuple form to describe the method: one conflict is one 4-tuple ,express and exist The nodes are occupied at the same time ;

[0051] Step 6.4: When a conflict is detected, activate the constraint tree branch generation mechanism: for the conflicting tuples , it will expand to generate two child nodes Nc1 and Nc2, and the two child nodes inherit all the constraints and solutions of the parent node;

[0052] Step 6.5: Add a new constraint to node Nc1 , add a new constraint for Nc2 ;

[0053] Step 6.6: Perform local path planning for the low-priority ship-brushing robot based on the constraints, while keeping the original paths of other ship-brushing robots unchanged.

[0054] Beneficial technical effects brought about by the present invention:

[0055] The present invention will The algorithm is combined with the Q-Learning algorithm to The algorithm is prone to fall into the situation of local optimal solution during the solution process. The Q-Learning algorithm is used to drive the iterative update of the Q table through multiple reward functions. During the algorithm integration phase, the Q value is mapped to a dynamic heuristic function H(n), significantly improving the algorithm's search efficiency and obstacle avoidance capabilities in complex environments. In terms of collaborative path planning, by creating a robot priority mechanism, the strategy divides multiple boat-brushing robots into different priorities and adds a binary constraint tree. When a conflict is detected between different robot paths, two child nodes are generated for each conflict and constraints are added. The affected path is replanned for the robot with the lower priority, avoiding the mutual influence of multiple robots during collaborative operations. The technical solution of the present invention can improve the collaborative work efficiency of multiple boat-brushing robots and solve the technical problems of dynamic obstacle avoidance response lag and low efficiency of multi-machine collaboration in traditional path planning under complex curved surface environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flowchart of basic information preprocessing in the present invention.

[0057] Figure 2 The improved The algorithm generates an initial path flow chart for the ship-washing robot group.

[0058] Figure 3 This is a flow chart for outputting the optimal path solution in the present invention. DETAILED DESCRIPTION

[0059] The specific implementation of the present invention will be further described below with reference to specific embodiments:

[0060] A method for collaborative operation path planning of multiple brush boat robots includes the following steps:

[0061] Step 1: Initialize the parameters, including: hull surface map information, terrain obstacles and constraints, and the starting and ending point information of the ship washing robot;

[0062] Step 2: Collect real-time data from sensors on the ship-washing robot, including encoders, IMUs, and temperature and humidity sensors. The encoders collect the real-time speed of the ship-washing robot while it is working, the IMUs collect the position coordinates and posture information of the ship-washing robot, and the temperature and humidity sensors monitor changes in the temperature and humidity of the seawater.

[0063] The encoder monitors the speed and displacement of the robot's drive wheels, providing feedback on its motion status through real-time speed data (such as linear velocity and angular velocity). Combined with pre-set path planning, it corrects speed deviations in real time to reduce path tracking errors. The IMU provides three-dimensional attitude angles (pitch, roll, and yaw) and acceleration information, building a real-time six-degree-of-freedom model of the robot's position. When the robot operates on complex curved surfaces, the contact pressure of the cleaning arm is dynamically adjusted based on this attitude data to avoid cleaning omissions or equipment wear due to surface unevenness. Temperature and humidity sensors monitor seawater temperature, air humidity, and surface condensation, reflecting real-time changes in the robot's working environment. By integrating multi-sensor data, a redundant perception system is constructed to reduce the risk of failure caused by a single sensor failure.

[0064] Step 3: Determine the number of ship washing robots and allocate the same type of robots according to the task scale, numbered as 、 、 etc. represent different individuals, formulate robot group priority strategy, and follow the order of numbers (such as > > ) assign priorities to ensure that there are no conflicts in the paths when multiple robots collaborate; in the priority planning framework, the path of the high-priority robot is locked after successful path planning, and its planned path information is updated before the next robot's path planning to avoid collisions.

[0065] Step 4: Use the grid contour method to rasterize the hull surface. The specific content is: obtain the geometric data of the hull surface, select one of the geometric reference points as the origin, determine the grid size according to the cleaning range of the robot, and generate a grid covering the entire hull surface area to be cleaned, such as Figure 1 As shown;

[0066] Specifically, the two-dimensional boundary coordinates of the hull surface to be cleaned are first obtained. The hull's outline is defined by a series of points, forming a closed polygon. The mesh resolution, as well as the number of rows and columns required to cover the entire hull, are then determined. Based on the set parameters, a mesh is created that covers the entire hull area. By discretizing the continuous surface into grid cells, the complex geometric problem is simplified into a discrete topological problem. Combined with a priority strategy to dynamically allocate tasks, this reduces the dimensionality of the path search and significantly shortens computation time.

[0067] Step 5: Use the improved The algorithm generates the initial path for the ship-washing robot group;

[0068] The algorithm is implemented using two lists: open and closed. The open list contains a list of nodes to be checked, where the path may pass through the nodes stored in it. The closed list contains nodes that have already been visited. First, the nodes adjacent to the starting point and that can be visited are stored in the open list. They are sorted in the list from smallest to largest according to their evaluation value. Since the node with the smallest evaluation value is the first node, only the first node in the list is the node to be visited. Then, if its adjacent node is the best node, it is saved in the closed list, and the adjacent optional nodes that have this node as a parent node are added to the open list. The selection continues until the target point is reached.

[0069] The robot's working area is divided by the grid method, and the working space is simplified into a two-dimensional grid. Each grid is numbered by the grid array to save the environmental information. The environmental information can represent the starting point, target point, obstacles and free area. The algorithm starts from the starting point A and checks its adjacent nodes in turn, and then continues to expand outward until it searches for the target point B. The grid passed from point A to point B is the path between AB.

[0070] The correct selection of the evaluation function will directly affect Whether the algorithm is successful or not, the determination of the function is closely related to the actual situation, so the choice of heuristic function is the key. An inappropriate heuristic function will lead to The quality of the algorithm's path planning decreases. The closer the estimated value is to the actual value, the more appropriate the heuristic function is. In order to prevent the algorithm from converging to a suboptimal solution too early during the search process and thus falling into a local optimum, the Q-Learning algorithm is used to optimize the heuristic function H(n). The Q-Learning algorithm is based on the Bellman equation and iteratively updates h(n)=max based on the immediate reward after action selection and the optimal Q value of the next state. a Q(s,a), so that it gradually approaches the minimum cumulative cost from the current state to the target. At the same time, the ε-greedy strategy is used to retain the exploration of unknown states, so that Reduce invalid node expansion in algorithms such as [1] and [2], and improve search efficiency and path quality.

[0071] like Figure 2 As shown, step 5 includes the following sub-steps:

[0072] Step 5.1, initialize the generated sets Open List and Close List. The Open List stores all nodes to be expanded, and the Close List stores all expanded nodes. Insert the start node into the Open List.

[0073] Step 5.2: When the Open List is not empty, select the node with the lowest total cost from the Open List. The total cost of the nodes The calculation formula is:

[0074] ;

[0075] in, is the actual path value from the starting point to the current point, To estimate the path cost from the current point to the target point;

[0076] Optimize using Q-Learning algorithm , Commonly used Euclidean distance formula, Manhattan distance formula and Chebyshev distance formula, the present invention adopts the Euclidean distance formula, The expression is:

[0077] ;

[0078] in, is the coordinate of the current node; is the coordinate of the starting node;

[0079] The node with the lowest total cost is removed from the Open List, added to the Close List, and used as the current node;

[0080] Step 5.3: If the current node is the target node, backtrack to the parent node to generate a path and end the algorithm. If the current node is not the target node, proceed to the next step.

[0081] Step 5.4. Find all the adjacent nodes of the current node. For each adjacent node, first determine whether it is in the Close List. If it is in the Close List, skip the node. If not, calculate the total cost of the adjacent nodes. ; Secondly, calculate the total cost of the adjacent nodes; then determine whether it is in the Open List. If not, add it to the Open List and update the total cost of the adjacent nodes and the parent node. If it is, compare the current total cost and the total cost calculated previously , if the current total cost is smaller, then update the parent node and update the total cost of the adjacent nodes to ;

[0082] Step 5.5: Return to step 5.2 and continue the loop until the target node is found.

[0083] In step 5.2, the total cost of the node needs to be calculated , where the heuristic function When the acceptability criterion (i.e., the cost is always no higher than the actual path cost) and the consistency condition is met, the algorithm can effectively guarantee the generation of the global optimal path. However, in actual engineering applications, if the parameters of the heuristic function are improperly set or there is a deviation in the heuristic guidance direction, the search process may converge to a suboptimal solution prematurely, which is manifested as the algorithm redundantly expanding in a specific area and failing to break through the constraints of the local extreme value, thus falling into the local optimum, thereby affecting the quality and efficiency of the overall path planning. The Q-Learning algorithm is used to optimize the heuristic function H(n). The Q-Learning algorithm is based on the Bellman equation and iteratively updates the immediate reward after action selection and the optimal Q value of the next state. , so that it gradually approaches the minimum cumulative cost from the current state to the goal. At the same time, the ε-greedy strategy is used to retain the exploration of unknown states, so that In algorithms such as , the expansion of invalid nodes is reduced, and the search efficiency and path quality are improved. The specific process includes the following steps:

[0084] Step 5.2.1: Initialize the Q-Learning algorithm parameters, obstacle information in the environment, number of robots, and coordinates of the starting and target points;

[0085] Construct the basic environment model for robot path planning;

[0086] Discretize the working area into uniform grids, and mark each grid as free space, obstacle or task target point;

[0087] The robot state space is defined as the two-dimensional coordinate (x, y) and the distribution of obstacles in its surrounding 3×3 neighborhood;

[0088] Define the robot's action space , The robot moves from the current grid unit to the adjacent Move one step in each direction, corresponding to the set of eight neighborhood movement directions (orthogonal directions: N, S, E, W; diagonal directions: NE, NW, SE, SW). Feasibility verification must be performed before each action is executed: the target grid must be within the map boundary and not occupied by static obstacles;

[0089] Step 5.2.2: Use the R-Learning algorithm to obtain the optimal strategy by maximizing the reward value and introducing a collision avoidance reward and penalty mechanism. Specifically:

[0090] The total reward obtained by the boat-cleaning robot in the current state s after executing action a after selecting the strategy is composed of the collision avoidance reward and the step reward, which can be expressed as:

[0091] ;

[0092] in, It is a sparse reward function that uses a target reward to directly motivate the path endpoint to converge, and a collision penalty to avoid path collisions. Usually, Q-Learning selects a fixed value as the reward function. The reward value for the action reaching the endpoint is 100, the reward value for the action without collision is -1, and the reward value for the action with collision is -50. The expression is:

[0093] ;

[0094] The distance reward function is used in the long-distance trajectory planning scenario of a ship-cleaning robot. The algorithm lacks an effective guidance mechanism in the initial phase, resulting in significant randomness in the exploration process, a large number of invalid iterations, and high computational resource consumption. To optimize algorithm performance, a distance reward and penalty mechanism is introduced to improve the reward function of the Q-Learning algorithm. This mechanism establishes the following evaluation function through mathematical modeling:

[0095] ;

[0096] Where: is the reward coefficient, The distance reward is the Euclidean distance between the starting point and the end point of the target. The distance reward can reduce the calculation time of the Q-Learning algorithm, eliminate redundant paths, and optimize path efficiency. In the initial exploration of the ship-farming robot, it tends to choose actions close to the target point.

[0097] During the training process, the Q value update follows the Bellman optimal equation, which is expressed as:

[0098] ;

[0099] Where, 、 is the state at the tth iteration and the t+1th iteration; 、 are the actions at the tth iteration and t+1th iteration respectively, and are learning rate and discount factor, respectively, with a value range of [0,1], For particles in state Next action The instant rewards you receive, The particle is in the state Take action The expected Q value of is the maximum expected future reward for all possible actions in the new state;

[0100] Step 5.2.3: Get the Q value of each state-action pair through Q-Learning training, traverse all grid nodes, and record the The maximum Q value of the grid nodes , store the maximum Q value result in the matrix, the matrix index is the node coordinate, the value is the maximum Q value corresponding to the node, and the trained Q value is embedded in Algorithm, reconstruction heuristic function H(n), its formula is:

[0101] ;

[0102] Where: Starting from a node, the maximum Q value that can be obtained by selecting the optimal action is the Q value, which represents the long-term benefits in historical experience; is the Euclidean distance from the current node to the target node, which is used to dynamically adjust the weight according to the distance from the node to the target; parameter is the attenuation coefficient, and its value range is [0.3,1].

[0103] Step 6: Use a binary tree constraint tree to detect whether there are conflicts in the solutions of each node in the initial path of the ship-cleaning robot population; if a conflict is detected in the current path, add constraints to two child nodes for each conflict and replan the affected path for the robot with low priority; the algorithm terminates when all paths are conflict-free and the total cost is optimal.

[0104] The mechanism achieves conflict resolution and global optimization of group paths through a three-stage progressive process: first, a spatiotemporal conflict detection model is established for the initial path set, and a quadtree spatial index and time window prediction algorithm are used to identify resource competition problems in the node solution space; second, a bidirectional constraint branch is constructed for the detected conflict events - each conflict point generates left and right child nodes with a logical inheritance relationship. The left branch inherits the parent's physical constraints and superimposes dynamic priority rules, while the right branch inherits the timing constraints and expands the safety buffer zone; finally, a priority-oriented path correction strategy is implemented to re-plan the paths of robots with lower priorities in the conflict until all paths are conflict-free and the total cost is optimal, at which time the algorithm terminates.

[0105] like Figure 3 As shown, step 6 includes the following sub-steps:

[0106] Step 6.1: Create a binary constraint tree based on the conflict search mechanism. Each constraint tree node N contains three pieces of information: the constraint set, the solution to the problem, and the cost of solving the problem at that node. The constraint set contains constraints for all the ship-farming robots in the problem.

[0107] Step 6.2: Algorithm initialization phase is improved The algorithm generates an unconstrained optimal path for each boat-cleaning robot to form an initial solution set;

[0108] Step 6.3: Start the conflict detection mechanism, using the 4-tuple form to describe the method: one conflict is one 4-tuple ,express and exist The nodes are occupied at the same time ;

[0109] Step 6.4: When a conflict is detected, activate the constraint tree branch generation mechanism: for the conflicting tuples , it will expand to generate two child nodes Nc1 and Nc2, and the two child nodes inherit all the constraints and solutions of the parent node;

[0110] Step 6.5: Add a new constraint to node Nc1 , add a new constraint for Nc2 ;

[0111] Step 6.6: Perform local path planning for the low-priority ship-brushing robot based on the constraints, while keeping the original paths of other ship-brushing robots unchanged.

[0112] Step 7: Smoothly optimize the path, traverse all nodes on the robot path, and when there are no obstacles on the line connecting the front and rear nodes of a node, delete the redundant nodes of the path, retaining only the starting and ending points and inflection points, and then delete the redundant inflection points to extract the key nodes as the intermediate target points, and output the optimal path plan.

[0113] The generated path is smoothed and optimized. Path optimization is an important post-processing step in mobile robot motion planning. Initial paths generated by algorithms often contain redundant turning points and polyline connections, causing the trajectory to deviate from optimal kinematic characteristics. To improve path continuity and practicability, this study proposes a path smoothing strategy based on node optimization. Specifically, the path node sequence is first traversed and analyzed. If the lines connecting adjacent nodes do not intersect the obstacle area, the intermediate node is removed. By iteratively executing this optimization process, unnecessary turning points in the path are effectively eliminated, ultimately generating a smooth trajectory that satisfies kinematic constraints.

[0114] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A method for collaborative operation path planning of multiple brush boat robots, characterized in that: The following steps are involved: Step 1: Initialize the parameters, including: hull surface map information, terrain obstacles and constraints, and the starting and ending point information of the ship washing robot; Step 2: Collect real-time data from sensors on the ship-washing robot, including encoders, IMUs, and temperature and humidity sensors. The encoders collect the real-time speed of the ship-washing robot while it is working, the IMUs collect the position coordinates and posture information of the ship-washing robot, and the temperature and humidity sensors monitor changes in the temperature and humidity of the seawater. Step 3: Determine the number of ship-washing robots, allocate the same type of robots according to the task size, and formulate a robot group priority strategy to ensure that there are no conflicts in the paths when multiple robots collaborate; Step 4: Use the grid contour method to perform grid processing on the hull surface; Step 5: Use the improved The algorithm generates the initial path for the ship-washing robot group; Step 6: Add a binary tree constraint tree to detect whether there are conflicts among the solutions of each node in the initial path of the boat-cleaning robot population. If a conflict is detected in the current path, add constraints to each of the two child nodes generated for each conflict, and replan the affected path for the robot with low priority. The algorithm terminates when all paths are conflict-free and the total cost is optimal. Step 7: Smoothly optimize the path, traverse all nodes on the robot path, and when there are no obstacles on the line connecting the previous and next nodes of a certain node, delete the redundant nodes of the path, leaving only the starting and ending points and the inflection points. Then, delete the redundant inflection points, extract the key nodes as the intermediate target points, and output the optimal path solution. The step 5 includes the following sub-steps: Step 5.1, initialize the generated sets Open List and Close List. The Open List stores all nodes to be expanded, and the Close List stores all expanded nodes. Insert the start node into the Open List. Step 5.2: When the Open List is not empty, select the node with the lowest total cost from the Open List. The total cost of the nodes The calculation formula is: ； in, is the actual path cost from the starting point to the current point, is the estimated path cost from the current point to the target point; Q-Learning algorithm is used to optimize , Using the Euclidean distance formula, The expression is: ； in, is the coordinate of the current node; is the coordinate of the starting node; The node with the lowest total cost is removed from the Open List, added to the Close List, and used as the current node; Step 5.3: If the current node is the target node, backtrack to the parent node to generate a path and end the algorithm. If the current node is not the target node, proceed to the next step. Step 5.4, find all the adjacent nodes of the current node. For each adjacent node, first determine whether it is in the CloseList. If it is in the Close List, skip the node. If not, calculate the total cost of the adjacent nodes. ; Then determine whether it is in the Open List. If not, add it to the Open List and update the total cost of the adjacent nodes and the parent node. If it is, compare the current total cost and the total cost calculated previously , if the current total cost is smaller, then update the parent node and update the total cost of the adjacent nodes to ; Step 5.5: Return to step 5.2 and continue the loop until the target node is found. In step 5.2, the Q-Learning algorithm is used to optimize , including the following sub-steps: Step 5.2.1: Initialize the Q-Learning algorithm parameters, obstacle information in the environment, number of robots, and coordinates of the starting and target points; Construct the basic environment model for robot path planning; Discretize the working area into uniform grids, and mark each grid as free space, obstacle or task target point; The robot state space is defined as the two-dimensional coordinate (x, y) and the distribution of obstacles in its surrounding 3×3 neighborhood; Define the robot's action space , The robot moves from the current grid unit to the adjacent Each action moves one step in each direction, corresponding to a set of eight neighborhood movement directions. Before each action is executed, feasibility verification must be performed: the target grid must be within the map boundary and not occupied by static obstacles; Step 5.2.2: Use the R-Learning algorithm to obtain the optimal strategy by maximizing the reward value and introducing a collision avoidance reward and penalty mechanism. Specifically: The total reward obtained by the boat-cleaning robot in the current state s after executing action a after selecting the strategy is composed of the collision avoidance reward and the step reward, which can be expressed as: ； in, is a sparse reward function, expressed as: ； is the distance reward function, expressed as: ； Where: is the reward coefficient, is the Euclidean distance between the starting point and the end point of the target; During the training process, the Q value update follows the Bellman optimal equation, which is expressed as: ； Where, 、 is the state at the tth iteration and the t+1th iteration; 、 are the actions at the tth iteration and t+1th iteration respectively, and are learning rate and discount factor, respectively, with a value range of [0,1], For particles in state Next action The instant rewards you receive, The particle is in the state Take action The expected Q value of is the maximum expected future reward for all possible actions in the new state; Step 5.2.3: Get the Q value of each state-action pair through Q-Learning training, traverse all grid nodes, and record the The maximum Q value of the grid nodes , store the maximum Q value result into the matrix, the matrix index is the node coordinate, the value is the maximum Q value corresponding to the node; embed the trained Q value into Algorithm, reconstruction heuristic function , the formula is: ； Where: Starting from a node, the maximum Q value that can be obtained by selecting the optimal action is the Q value, which represents the long-term benefits in historical experience; is the Euclidean distance from the current node to the target node, which is used to dynamically adjust the weight according to the distance from the node to the target; parameter is the attenuation coefficient, and its value range is [0.3,1]; Step 6 includes the following sub-steps: Step 6.1: Create a binary constraint tree based on the conflict search mechanism. Each constraint tree node N contains three pieces of information: the constraint set, the solution to the problem, and the cost of solving the problem at that node. The constraint set contains constraints for all the ship-farming robots in the problem. Step 6.2: Algorithm initialization phase is improved The algorithm generates an unconstrained optimal path for each boat-cleaning robot to form an initial solution set; Step 6.3: Start the conflict detection mechanism, using the 4-tuple form to describe the method: one conflict is one 4-tuple ,express and exist The nodes are occupied at the same time ; Step 6.4: When a conflict is detected, activate the constraint tree branch generation mechanism: for the conflicting tuples , it will expand to generate two child nodes Nc1 and Nc2, and the two child nodes inherit all the constraints and solutions of the parent node; Step 6.5: Add a new constraint to node Nc1 , add a new constraint for Nc2 ; Step 6.6: Perform local path planning for the low-priority ship-brushing robot based on the constraints, while keeping the original paths of other ship-brushing robots unchanged.

Images