Method and system for optimizing robot path planning based on deep learning

Abstract

The invention discloses a method and system for optimizing robot path planning based on deep learning, and the method comprises the steps: building a model for a path planning problem of a robot, converting the path planning problem into an optimizable mathematical problem, defining the mathematical problem by a target function of an optimization task, and enabling the target function to comprise length cost and collision cost; encoding environmental information in robot motion planning by a BPS that captures features of the environment by a set of predefined base points and generates feature vectors by calculating the distance of these points to the nearest obstacle in the environment; based on the generated feature vector, completing model training through deep learning, and performing path planning model training by using a neural network to obtain an optimal path; the path planning model is trained based on deep learning to obtain the optimal path, and the optimal path is used as the initial hot start of the path planning algorithm based on optimization, so that the path planning model training is more efficient, and the planning efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of robot path planning, and in particular to a method and system for optimizing robot path planning based on deep learning. Background Art

[0002] The goal of robot motion planning is to find a collision-free path from an initial joint configuration to a target configuration safely and efficiently in a given environment. In this process, the robot needs to consider various factors, including path length, motion time, and possible obstacles.

[0003] The core of motion planning is to solve two main problems: one is how to find a feasible path in a complex environment; the other is how to ensure that the robot does not collide with obstacles or itself during movement. In order to solve these problems, robot motion planning usually relies on accurate environment modeling, efficient path search algorithms, and real-time collision detection mechanisms. The current mainstream robot motion planning is mainly based on sampling motion planning and optimization-based motion planning methods.

[0004] Sampling-based motion planning methods, mainly the rapidly expanding random tree (RRT) and probabilistic roadmap (PRM) methods and their variants, construct the robot's motion trajectory by randomly sampling in the robot's workspace. This method does not require precise modeling of the environment and is suitable for complex and uncertain environments.

[0005] However, the inventors of this application found that the above technology has at least the following technical problems in the process of implementing the technical solution of the invention in the embodiment of this application: Optimization-based motion planning methods focus more on finding the optimal solution for specific performance indicators. Such methods usually transform the motion planning problem into an optimization problem, and then use optimization algorithms such as gradient descent and genetic algorithms to solve it. However, since the objective function of robot motion planning often contains multiple different local minima, it is easy for the optimization algorithm to converge prematurely, making it impossible to find the global optimal solution. In order to solve this problem, current solutions mostly use the multi-start method, that is, starting the optimization search from multiple different initial points to increase the possibility of finding the global optimal solution that meets the performance indicators. The computational efficiency of the sampling-based method will be affected by the scale of the problem. When faced with complex multi-robot motion planning problems, it is easy to experience an explosion of computational dimensions. Summary of the invention

[0006] The embodiments of the present application provide a method and system for optimizing robot path planning based on deep learning, thereby solving the problem that the traditional sampling-based methods in the prior art have low computational efficiency and are prone to computational dimension explosion when facing complex multi-robot motion planning problems. A neural network suitable for robot path planning is proposed to encode the optimal motion trajectory in different tasks and environments; a basic point set is introduced into the motion planning problem based on optimization to make the path planning model training more efficient; and path planning using deep learning prediction is proposed as a hot start method for the optimization-based path planning solution, which greatly improves the planning efficiency.

[0007] The embodiment of the present application provides a method for optimizing robot path planning based on deep learning, comprising: S1. Modeling the path planning problem of the robot and converting it into a mathematical problem that can be optimized, wherein the mathematical problem is defined by an objective function of the optimization task, wherein the objective function includes a length cost and a collision cost; S2, encoding the environment information in the robot motion planning through BPS, wherein the BPS captures the characteristics of the environment through a set of predefined basic points and generates a feature vector by calculating the distance from these points to the nearest obstacle in the environment; S3. Based on the generated feature vector, complete the model training through deep learning, and use the neural network to train the path planning model to obtain the optimal path; S4. Based on deep learning, the path planning model is trained to obtain the optimal path, which serves as the initial hot start of the optimization-based path planning algorithm, thereby connecting to any optimization-based path planning algorithm.

[0008] Furthermore, in step S1, the objective function of the optimization task specifically includes: The objective function of the optimization task is defined as: ; Q* is the objective function of the optimization task. This formula defines the ultimate goal of motion planning, which is to find a path Q so that the objective function U(Q) reaches the minimum value; the objective function U(Q) is a combination of the length cost and the collision cost.

[0009] Furthermore, in step S2, the main steps of BPS encoding specifically include: S21. Based on the environmental information in the robot motion planning, select a set of predefined basic point sets ; in, is a fixed position in the environment. For each base point, find the distance to the nearest obstacle in the environment. S22, combining the distance from each basic point to the nearest obstacle into a feature vector, the environmental information is converted into a feature vector of a fixed size, and the feature vector is expressed by the formula: ; in Indicates The location of an obstacle.

[0010] Furthermore, in step S3, the path planning model training specifically includes: S31, creating an initial data set G including multiple motion tasks and their corresponding optimal paths based on the optimized motion planner; S32, based on the initial data set G of the optimal path, a feedforward neural network is constructed using multiple tapered fully connected layers, the input layer receives the environmental features encoded by the basic point set, and the starting position and configuration of the robot, and the output result is the node of the predicted optimal path; S33, use the initialization data set G to train the neural network model ; S34, when the performance of the neural network on the test set improves, it enters the iteration process; S35, in each iteration, retrain the neural network model using the improved dataset .

[0011] Furthermore, in step S4, it specifically includes: The main steps of path planning after hot start are as follows: The initial path of the OMP is predicted through the pre-trained path planning deep learning model; Based on the initial path of OMP, run the OMP algorithm and try to improve the initial path through iterative optimization; After each path optimization, the collision between the path and environmental obstacles is checked. If a collision occurs, the path is replanned. The additional paths and nodes added by the replanning will also be predicted by the model. When the path cost reaches the specified requirement, or the path cost no longer decreases after a certain number of iterations, the algorithm stops and returns the optimal path.

[0012] A system for optimizing robot path planning based on deep learning, comprising: A task definition and initialization module, which is used to model the path planning problem of the robot and convert it into a mathematical problem that can be optimized. The mathematical problem is defined by the objective function of the optimization task, and the objective function includes length cost and collision cost; A BPS encoding module, for encoding environmental information in robot motion planning through BPS, wherein the BPS captures the characteristics of the environment through a set of predefined basic points and generates a feature vector by calculating the distances from these points to the nearest obstacles in the environment; The path planning model training module is used to complete the model training through deep learning based on the generated feature vectors. The path planning model training is performed using a neural network to obtain the optimal path. The initial hot start module is based on deep learning to train the path planning model to obtain the optimal path, which serves as the initial hot start of the optimization-based path planning algorithm, thereby connecting to any optimization-based path planning algorithm.

[0013] Furthermore, in the task definition and initialization module, include, The objective function of the optimization task is defined as: ; Q* is the objective function of the optimization task. This formula defines the ultimate goal of motion planning, which is to find a path Q such that the objective function Reach the minimum value; objective function is a combination of the length cost and the collision cost.

[0014] Furthermore, in the BPS encoding module, it includes: Basic point set pre-determined unit: based on the environment information in the robot motion planning, select a set of pre-defined basic point sets ; in, is a fixed position in the environment. For each base point, find the distance to the nearest obstacle in the environment. Feature vector unit: The distance from each base point to the nearest obstacle is combined into a feature vector. The environmental information is converted into a feature vector of a fixed size. The feature vector is expressed as follows: ; in Indicates The location of an obstacle.

[0015] Furthermore, in the path planning model training module, it includes: A data set creation unit, used for creating an initial data set G including a plurality of motion tasks and their corresponding optimal paths based on an optimized motion planner; The node prediction unit of the optimal path is used to construct a feedforward neural network using multiple tapered fully connected layers based on the initial data set G of the optimal path. The input layer receives the environmental features encoded by the basic point set and the starting position and configuration of the robot. The output result is the predicted node of the optimal path. The neural network model training unit is used to train the neural network model using the initialization data set G. ; Iteration unit, used to enter the iteration process when the performance of the neural network on the test set improves; The retraining unit is used to retrain the neural network model using the improved dataset in each iteration. .

[0016] Furthermore, in the initial hot start module, it includes: The main steps of path planning after hot start are as follows: The initial path of the OMP is predicted through the pre-trained path planning deep learning model; Based on the initial path of OMP, run the OMP algorithm and try to improve the initial path through iterative optimization; After each path optimization, the collision between the path and environmental obstacles is checked. If a collision occurs, the path is replanned. The additional paths and nodes added by the replanning will also be predicted by the model. When the path cost reaches the specified requirement, or the path cost no longer decreases after a certain number of iterations, the algorithm stops and returns the optimal path.

[0017] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. A neural network suitable for robot path planning is proposed to encode the optimal motion trajectory in different tasks and environments; 2. Introducing the basic point set into the optimization-based motion planning problem to make the path planning model training more efficient; 3. It is proposed to use deep learning predicted path planning as a hot start method for the optimization-based path planning scheme, which greatly improves the planning efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A flowchart of a method and system for optimizing robot path planning based on deep learning; Figure 2 Block diagram of the method and system for optimizing robot path planning based on deep learning. DETAILED DESCRIPTION

[0019] A robot path planning optimization method based on deep learning, the main module of the method is as follows Figure 1 The trivia includes a task definition and initialization module, a BPS encoding module, a path planning model training module, and a complete motion path planning solution built with these modules based on the optimized motion planning (OMP) method. The BPS is a new method designed to simplify and accelerate point cloud data learning.

[0020] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0021] The embodiment of the present application provides a method for optimizing robot path planning based on deep learning, comprising: S1. Modeling the path planning problem of the robot and converting it into a mathematical problem that can be optimized, wherein the mathematical problem is defined by an objective function of the optimization task, wherein the objective function includes a length cost and a collision cost; S2, encoding the environment information in the robot motion planning through BPS, wherein the BPS captures the characteristics of the environment through a set of predefined basic points and generates a feature vector by calculating the distance from these points to the nearest obstacle in the environment; S3. Based on the generated feature vector, complete the model training through deep learning, and use the neural network to train the path planning model to obtain the optimal path; S4. Based on deep learning, the path planning model is trained to obtain the optimal path, which serves as the initial hot start of the optimization-based path planning algorithm, thereby connecting to any optimization-based path planning algorithm.

[0022] Furthermore, in step S1, the objective function of the optimization task specifically includes: The objective function of the optimization task is defined as: ; Q* is the objective function of the optimization task. This formula defines the ultimate goal of motion planning, which is to find a path , so that the objective function Reach the minimum value; objective function is a combination of the length cost and the collision cost.

[0023] Furthermore, in step S2, the main steps of BPS encoding specifically include: S21, based on the environmental information in the robot motion planning, selecting a set of predefined basic point sets; ; in, is a fixed position in the environment. For each base point, find the distance to the nearest obstacle in the environment. S22, combining the distance from each basic point to the nearest obstacle into a feature vector, the environmental information is converted into a feature vector of a fixed size, and the feature vector is expressed by the formula: ; in Indicates The location of an obstacle.

[0024] Furthermore, in step S3, the path planning model training specifically includes: S31, creating an initial data set G including multiple motion tasks and their corresponding optimal paths based on the optimized motion planner; S32, based on the initial data set G of the optimal path, a feedforward neural network is constructed using multiple tapered fully connected layers, the input layer receives the environmental features encoded by the basic point set, and the starting position and configuration of the robot, and the output result is the node of the predicted optimal path; S33, use the initialization data set G to train the neural network model ; S34, when the performance of the neural network on the test set improves, it enters the iteration process; S35, in each iteration, retrain the neural network model using the improved dataset .

[0025] Furthermore, in step S4, it specifically includes: The main steps of path planning after hot start are as follows: The initial path of the OMP is predicted through the pre-trained path planning deep learning model; Based on the initial path of OMP, run the OMP algorithm and try to improve the initial path through iterative optimization; After each path optimization, the collision between the path and environmental obstacles is checked. If a collision occurs, the path is replanned. The additional paths and nodes added by the replanning will also be predicted by the model. When the path cost reaches the specified requirement, or the path cost no longer decreases after a certain number of iterations, the algorithm stops and returns the optimal path.

[0026] A system for optimizing robot path planning based on deep learning, comprising: Task definition and initialization module 01, used to model the path planning problem of the robot and convert it into a mathematical problem that can be optimized. The mathematical problem is defined by the objective function of the optimization task, and the objective function includes length cost and collision cost; Specifically, the task definition and initialization module first models the problem and converts it into a mathematical problem that can be optimized, which involves modeling the robot itself, its environment, obstacles, etc.

[0027] The motion planning task is defined as starting from To the target The path planning problem is to avoid collisions with obstacles in the environment and the robot itself. The path Q is represented as a collection of nodes: ; in, Indicates that at the node Position is the key configuration of the robot.

[0028] Objective Function It represents the cost of the robot moving to a specified position. The objective function varies depending on the task. In most task scenarios, the objective function includes the length cost. and collision costs : ; Among them, the length cost It is expressed in the following form: ; Represents the total number of nodes on path Q.

[0029] Collision Cost It is expressed in the following form: ; Among them, N f Represents the number of rigid bodies of the robot, such as the robot's links and joints; N u represents the u-th node; N si Represents the number of shapes on the i-th rigid body; t,u means summing t and u means facilitating the middle points between nodes u and t on the path, and then performing double summation for each middle point; This formula calculates the cost of a robot colliding with obstacles in the environment during movement, taking into account all rigid bodies of the robot and all geometric shapes on each rigid body, such as the distance D between a sphere and obstacles in the environment, and using function c to calculate the collision cost. represents two adjacent nodes, represents the forward kinematics function of the robot, representing the positions of the various parts of the machine, Indicated in Position The center point of the machine, Indicates Position The radius of the machine, represents the distance between the machine and the nearest obstacle, and C(d) represents the penalty function when a collision occurs, which is as follows: ; The function is used to handle collision detection in path planning. When the distance between a part of the robot and an obstacle is When it is a negative number (i.e. a collision occurred), the function returns Indicates the distance of the penalty path entering the obstacle. When the distance between the machine and the obstacle is greater than 0 and less than When , the function returns a This helps prevent the machine from moving too close to obstacles. No punishment, It is a constant set in advance according to the task.

[0030] The objective function of the optimization task is defined as follows: ; This formula defines the ultimate goal of motion planning, which is to find a path Q such that the objective function Reach the minimum value. Objective function It is a combination of length cost and collision cost, which reflects the length, smoothness and safety of the path. By minimizing this objective function, an efficient and collision-free motion path can be obtained.

[0031] A BPS encoding module 02, for encoding environmental information in robot motion planning through BPS, wherein the BPS captures the characteristics of the environment through a set of predefined basic points and generates a feature vector by calculating the distances from these points to the nearest obstacles in the environment; Specifically, BPS (Basis Point Set) encoding is an environment representation method of the present invention, which is used to efficiently encode the environment information in the robot motion planning, and then complete the model training. BPS captures the characteristics of the environment through a set of predefined basis points (basis points), and generates feature vectors by calculating the distances from these points to the nearest obstacles in the environment.

[0032] Basic point set pre-determining unit 05: based on the environment information in the robot motion planning, select a set of pre-defined basic point sets; ; in, is a fixed position in the environment. For each base point, find the distance to the nearest obstacle in the environment. Feature vector unit 06: The distance from each basic point to the nearest obstacle is combined into a feature vector. The environmental information is converted into a feature vector of a fixed size. The feature vector is expressed by the formula: ; in Indicates The environmental information is converted into a fixed-size feature vector through BPS encoding, which enables the neural network to effectively process and learn the relationship between environmental features and robot motion planning.

[0033] The path planning model training module 03 is used to complete the model training through deep learning based on the generated feature vector, and the optimal path is obtained by training the path planning model using a neural network; Specifically, it aims to improve and enhance the training dataset of a neural network by using it. The purpose of this process is to improve the prediction quality of the neural network in motion planning tasks and ensure that the network can work reliably in diverse and challenging environments.

[0034] Specifically, in the path planning model training module, it includes: A data set creation unit 07, used to create an initial data set G including a plurality of motion tasks and their corresponding optimal paths based on an optimized motion planner; Specifically, an initial data set G containing multiple motion tasks and their corresponding optimal paths is created using an optimization-based motion planner (OMP). These data sets include input (environment and start / target configuration) and output (optimal path). The motion planner here can be any motion planning algorithm. In the present invention, the CHOMP algorithm is selected, which is an optimization method that focuses on computational efficiency.

[0035] The node prediction unit 08 of the optimal path is used to construct a feedforward neural network using multiple tapered fully connected layers based on the initial data set G of the optimal path, the input layer receives the environmental features encoded by the basic point set, and the starting position and configuration of the robot, and the output result is the predicted node of the optimal path; Specifically, multiple conical fully connected layers are used to construct a feedforward neural network with a structure similar to DenseNet. Its input layer receives environmental features encoded by the Basis Point Set (BPS), as well as the robot's starting position and configuration, and the output result is the node of the predicted optimal path.

[0036] The model's loss function uses the mean squared error (MSE) to evaluate the difference between the path predicted by the network and the OMP optimal path: ; in Represents a series of nodes in the path predicted by the model, Table Optimal path nodes calculated by OMP algorithm.

[0037] Neural network model training unit 09, used to train the neural network model using the initialization data set G ; An iteration unit 10, used to enter an iteration process when the performance of the neural network on the test set is improved; First, clean the dataset: For each sample in the dataset, use the neural network model To predict a path , and use OMP to find a better path for this prediction .

[0038] Calculate the cost of two paths if Cost U(y) is less than The cost is recalculated using the multi-start OMP algorithm. ,when Costs are still higher than When using To replace the labels of the original samples, thereby improving the quality of the data set.

[0039] Expand the data set again: randomly generate new task samples , and use the neural network model To predict the path and calculate the cost ,when Less than the constant threshold set in advance for the task When , it means that the task is too simple and can be ignored. Otherwise, OMP is used to generate the path. , and added to the dataset, thereby expanding the dataset and increasing the proportion of challenging samples.

[0040] The retraining unit 11 is used to retrain the neural network model using the improved data set in each iteration. .

[0041] Specifically, in each iteration, the neural network model is retrained using the improved dataset. . In this way, the network continuously learns and adapts from the dataset to improve its prediction accuracy and generalization ability. This combination of adaptive dataset expansion and neural network training enables the neural network to perform better in solving complex motion planning tasks. After continuously cleaning and expanding the training dataset, the network is able to learn richer and more challenging motion planning results, thereby providing more reliable path planning solutions in practical applications.

[0042] The initial hot start module 04 is based on deep learning to train the path planning model, and is used to obtain the optimal path as the initial hot start of the optimization-based path planning algorithm, thereby connecting to any optimization-based path planning algorithm.

[0043] Specifically, the path planning optimization method based on deep learning proposed in the present invention uses the optimal path predicted by deep learning as the initial hot start of the optimization-based path planning algorithm, and can be connected to any optimization-based path planning algorithm. The main steps of path planning after hot start are as follows: 1) OMP path initialization The pre-trained path planning deep learning model is used to predict the initial path of OMP, which is more likely to be close to the true optimal path than random selection, which helps to improve the convergence speed and success rate of OMP.

[0044] 2) Path optimization Run the OMP algorithm to try to improve the initial path through iterative optimization. The goal of the optimization is to minimize the length, smoothness, and collision cost of the path. This process is repeated multiple times.

[0045] 3) Collision Detection After each path optimization, the collision between the path and environmental obstacles is checked. If a collision occurs, the path is replanned. The additional paths and nodes added by the replanning will also be predicted by the model.

[0046] 4) Convergence check When the path cost reaches the specified requirement, or the path cost no longer decreases after a certain number of iterations, the algorithm stops and returns the optimal path.

[0047] Through the above steps, the optimized path planning algorithm can effectively solve the collision conflict problem and find a safe and efficient path. Using the model predicted path as the optimization hot start can greatly reduce the time for the optimization algorithm to converge and improve planning efficiency.

[0048] Furthermore, in the task definition and initialization module, include, The objective function of the optimization task is defined as: ; This formula defines the ultimate goal of motion planning, which is to find a path , so that the objective function Reach the minimum value; objective function is a combination of the length cost and the collision cost.

[0049] Furthermore, in the BPS encoding module, it includes: Basic point set pre-determined unit: based on the environment information in the robot motion planning, select a set of pre-defined basic point sets ; in, is a fixed position in the environment. For each base point, find the distance to the nearest obstacle in the environment. Feature vector unit: The distance from each base point to the nearest obstacle is combined into a feature vector. The environmental information is converted into a feature vector of a fixed size. The feature vector is expressed as follows: ; in Indicates The location of an obstacle.

[0050] The present invention generates an optimization-based robot path planning solution through deep learning, which has the following advantages over traditional robot planning solutions: By using a deep learning network as a preprocessor, a well-informed initial solution is provided for the traditional Optimal Motion Planning (OMP), which greatly reduces the number of iterations required to find a feasible path. This is especially important in dynamic or time-sensitive application scenarios.

[0051] Using basis point sets (BPS) as a way to encode the environment allows neural networks to generalize in a variety of different contexts.

[0052] Traditional sampling-based methods have low computational efficiency and are prone to computational dimensionality explosion when faced with complex multi-robot motion planning problems.

[0053] The objective function of traditional optimization-based motion planning schemes often has many local minima, so a multi-start approach is required to find a feasible global solution, which makes the calculation time of optimization-based motion planning schemes long.

[0054] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0055] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0056] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0057] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0058] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0059] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for optimizing robot path planning based on deep learning, characterized in that: include, S1. Model the path planning problem of the robot and convert it into a mathematical problem that can be optimized. The mathematical problem is defined by the objective function of the optimization task, and the objective function includes length cost and collision cost. The objective function of the optimization task is defined as: ； Q* is the objective function of the optimization task. This formula defines the ultimate goal of motion planning, which is to find a path Q such that the objective function U(Q) reaches the minimum value. The objective function U(Q) is a combination of the length cost and the collision cost. S2, encoding the environment information in the robot motion planning through BPS, wherein the BPS captures the characteristics of the environment through a set of predefined basic points and generates a feature vector by calculating the distance from these points to the nearest obstacle in the environment; S3. Based on the generated feature vector, complete the model training through deep learning, and use the neural network to train the path planning model to obtain the optimal path; S4. Based on deep learning, the path planning model is trained to obtain the optimal path, which serves as the initial hot start of the optimization-based path planning algorithm, thereby connecting to any optimization-based path planning algorithm.

2. A method for optimizing robot path planning based on deep learning as claimed in claim 1, characterized in that: In step S2, the main steps of BPS encoding, which is a new method designed to simplify and accelerate point cloud data learning, specifically include: S21. Based on the environmental information in the robot motion planning, select a set of predefined basic point sets ； in, is a fixed position in the environment. For each base point, find the distance to the nearest obstacle in the environment. S22, combining the distance from each basic point to the nearest obstacle into a feature vector, the environmental information is converted into a feature vector of a fixed size, and the feature vector is expressed by the formula: ； in Indicates The location of an obstacle.

3. A method for optimizing robot path planning based on deep learning as claimed in claim 1, characterized in that: In step S3, the path planning model training specifically includes: S31, creating an initial data set G including multiple motion tasks and their corresponding optimal paths based on the optimized motion planner; S32, based on the initial data set G of the optimal path, a feedforward neural network is constructed using multiple tapered fully connected layers, the input layer receives the environmental features encoded by the basic point set, and the starting position and configuration of the robot, and the output result is the node of the predicted optimal path; S33, use the initialization data set G to train the neural network model ; S34, when the performance of the neural network on the test set improves, it enters the iteration process; S35, in each iteration, retrain the neural network model using the improved dataset .

4. A method for optimizing robot path planning based on deep learning as claimed in claim 1, characterized in that: In step S4, specifically including: The main steps of path planning after hot start are as follows: The initial path of the OMP is predicted through the pre-trained path planning deep learning model; Based on the initial path of OMP, run the OMP algorithm and try to improve the initial path through iterative optimization; After each path optimization, the collision between the path and environmental obstacles is checked. If a collision occurs, the path is replanned. The additional paths and nodes added by the replanning will also be predicted by the model. When the path cost reaches the specified requirement, or the path cost no longer decreases after a certain number of iterations, the algorithm stops and returns the optimal path.

5. A system for optimizing robot path planning based on deep learning, characterized in that: include, The task definition and initialization module is used to model the robot's path planning problem and convert it into a mathematical problem that can be optimized. The mathematical problem is defined by the objective function of the optimization task, which includes length cost and collision cost. The objective function of the optimization task is defined as: ； Q* is the objective function of the optimization task. This formula defines the ultimate goal of motion planning, which is to find a path Q such that the objective function U(Q) reaches the minimum value. The objective function U(Q) is a combination of the length cost and the collision cost. A BPS encoding module, for encoding environmental information in robot motion planning through BPS, wherein the BPS captures the characteristics of the environment through a set of predefined basic points and generates a feature vector by calculating the distances from these points to the nearest obstacles in the environment; The path planning model training module is used to complete the model training through deep learning based on the generated feature vectors. The path planning model training is performed using a neural network to obtain the optimal path. The initial hot start module is based on deep learning to train the path planning model to obtain the optimal path, which serves as the initial hot start of the optimization-based path planning algorithm, thereby connecting to any optimization-based path planning algorithm.

6. A system for optimizing robot path planning based on deep learning as claimed in claim 5, characterized in that: In the BPS coding module, including, Basic point set pre-determined unit: based on the environment information in the robot motion planning, select a set of pre-defined basic point sets ； in, is a fixed position in the environment. For each base point, find the distance to the nearest obstacle in the environment. Feature vector unit: The distance from each base point to the nearest obstacle is combined into a feature vector. The environmental information is converted into a feature vector of a fixed size. The feature vector is expressed as follows: ； in Indicates The location of an obstacle.

7. A system for optimizing robot path planning based on deep learning as claimed in claim 5, characterized in that: In the path planning model training module, including, A data set creation unit, used for creating an initial data set G including a plurality of motion tasks and their corresponding optimal paths based on an optimized motion planner; The node prediction unit of the optimal path is used to construct a feedforward neural network using multiple tapered fully connected layers based on the initial data set G of the optimal path. The input layer receives the environmental features encoded by the basic point set and the starting position and configuration of the robot. The output result is the predicted node of the optimal path. The neural network model training unit is used to train the neural network model using the initialization data set G ; Iteration unit, used to enter the iteration process when the performance of the neural network on the test set improves; The retraining unit is used to retrain the neural network model using the improved dataset in each iteration. .

8. A system for optimizing robot path planning based on deep learning as claimed in claim 5, characterized in that: In the initial hot start module, including, The main steps of path planning after hot start are as follows: The initial path of the OMP is predicted through the pre-trained path planning deep learning model; Based on the initial path of OMP, run the OMP algorithm and try to improve the initial path through iterative optimization; After each path optimization, the collision between the path and environmental obstacles is checked. If a collision occurs, the path is replanned. The additional paths and nodes added by the replanning will also be predicted by the model. When the path cost reaches the specified requirement, or the path cost no longer decreases after a certain number of iterations, the algorithm stops and returns the optimal path.

Images