A robot motion planning method inspired by human brain scene semantic cooperation mechanism

Abstract

The application provides a robot motion planning method and device inspired by human brain scene semantic cooperation mechanism, and relates to the technical field of robot motion planning.The method comprises the following steps: based on a robot intelligent operation platform, collecting data of a motion task of the robot to obtain a motion starting position, a motion target position, an obstacle position and map environment information; based on a perception path segment selector, predicting a path segment collision probability according to the motion starting position, the motion target position, the obstacle position and the map environment information to obtain a safe random geometric graph; according to the safe random geometric graph, performing heuristic value calculation through a global correction heuristic to obtain node heuristic values of the safe random geometric graph; and according to the node heuristic values and the safe random geometric graph, performing path search through a differentiable module to obtain a robot motion planning path.The application is a high-efficiency and flexible robot motion planning method combining a graph neural network planner and a scene semantic cooperation mechanism.

Description

Technical Field

[0001] This invention relates to the field of robot motion planning technology, and in particular to a robot motion planning method and apparatus inspired by the contextual semantic coordination mechanism of the human brain. Background Technology

[0002] Motion planning is a key technology in robotic systems, with significant application prospects in various fields such as transportation and medical systems. The goal of motion planning is to find a collision-free path connecting the starting point and the target point in the decision space, ideally minimizing the path cost. However, real-world decision spaces typically contain a large number of continuous variables and are high-dimensional unstructured spaces, which greatly increases the complexity of motion planning. Therefore, implementing fast and reliable motion planning methods in high-dimensional continuous spaces is crucial for improving its practical applications.

[0003] Classical motion planning solutions can be broadly categorized into two types: search-based methods and sampling-based methods. Search-based methods, such as A... Path search algorithms and Dijkstra's shortest path algorithm conceptualize the planning problem as a graph search problem. Sampling-based methods, including fast exploratory random trees and probabilistic route graphs, construct exploratory trees connecting start and destination points by randomly sampling in the configuration space. However, these classic methods are often inefficient when dealing with high-dimensional planning problems.

[0004] In high-dimensional continuous environments, learning-based planning methods are key to improving motion planning performance. Leveraging the powerful learning and representation capabilities of neural networks, these methods can extract key patterns from the configuration space of data or mimic the behavior of expert planners, optimizing crucial aspects of the planning process, such as sampling mechanisms, to achieve superior performance. Currently, various neural network-driven motion planners have been proposed, including methods based on convolutional neural networks, recurrent neural networks, and graph neural networks. Among these methods, graph neural network-based planners exhibit unique performance advantages in high-dimensional tasks. Graph neural network-based methods significantly improve planning efficiency in high-dimensional spaces by accessing random geometric graphs formed by sampling, eliminating the need to encode the entire workspace as in other learning-based methods. Furthermore, graph neural networks can directly manipulate graph-structured data, predicting the priority of nodes or edges in random geometric graphs through efficient graph pattern learning, reducing redundant computation, and improving the efficiency and quality of path exploration.

[0005] Among a series of graph neural network-based planners, the Graph Neural Network-based Explorer (GNN-Explorer) and the Graph Neural Network-based Motion Planning with Efficient Graph Search (GraphMP) are typical examples due to their high efficiency and high quality, but their potential has not yet been fully realized. For example, GNN-Explorer does not fully consider the impact of total path cost when exploring edge priorities, which to some extent limits the cost optimization of the final search path. To further reduce path costs, GraphMP introduces a checker to filter potential collision edges in the RGG and calculates the heuristic value of nodes based on the graph neural network, combined with A... The module searches for near-optimal paths on the processed random geometry, significantly improving planning efficiency and path quality.

[0006] However, existing graph neural network-based planning methods still have technical limitations when facing high-dimensional continuous environments. Traditional planning methods construct random geometric graphs based on uniform sampling strategies, that is, indiscriminately exploring the entire space, failing to fully utilize the structural information of the actual problem itself. Inappropriate edges and nodes may exist, and the ability to predict the priority of edges or nodes may be reduced by the influence of interfering edges, thus decreasing the global planning performance and reliability. Traditional planning methods often ignore the global topological structure, resulting in poor performance in sparse graphs, and they process each task independently, failing to fully utilize past experience and established data models. The encoding methods for nodes or edges mainly rely on local features and neighborhood-based message passing mechanisms. When the effective connections in the graph are significantly reduced, the graph representation capability of the graph neural network model is limited.

[0007] Humans exhibit unique advantages in decision-making and reasoning. Neuroscience research shows that the human ability to understand and plan for the surrounding world during decision-making relies on semantic and episodic memory systems. When faced with weak information that is difficult to retrieve, the global knowledge framework in semantic memory provides contextual support for episodic memory, helping humans make decisions with ambiguous cues; specific event details from episodic memory then complete or modify the semantic memory framework. This two-way collaborative mechanism is crucial for guiding decision-making, adapting to environmental changes, and enhancing cognitive flexibility.

[0008] In the existing technology, there is a lack of an efficient and flexible robot motion planning method that combines graph neural network planners and contextual semantic cooperation mechanisms. Summary of the Invention

[0009] To address the technical challenges of efficient planning and precise obstacle avoidance for robots in high-dimensional continuous environments without large-scale data and high computing power, this invention provides a robot motion planning method and apparatus inspired by the human brain's contextual semantic collaboration mechanism. The technical solution is as follows: On the one hand, a robot motion planning method inspired by the human brain's contextual semantic coordination mechanism is provided. This method is implemented by a robot motion planning device and includes: Based on the robot intelligent operation platform, data is collected on the robot's motion tasks to obtain the starting position of motion, the position of the target motion, the position of obstacles, and map environment information. Based on the perceptual road segment selector, the collision probability of the path segment is predicted according to the starting position of the movement, the position of the moving target, the position of the obstacle, and the map environment information, and a safe random geometry is obtained. Based on the secure random geometry graph, the node heuristic values ​​of the secure random geometry graph are obtained by calculating the heuristic values ​​through a global correction heuristic. Based on node heuristics and a secure random geometry graph, through differentiable A... The module performs path search to obtain the robot's motion planning path.

[0010] On the other hand, a robot motion planning device inspired by the human brain's contextual semantic collaboration mechanism is provided. This device is applied to a robot motion planning method inspired by the human brain's contextual semantic collaboration mechanism. The device includes: The information acquisition module is used to collect data on the robot's motion tasks based on the robot's intelligent operation platform, and obtain information such as the starting position of the motion, the position of the target motion, the position of obstacles, and the map environment. The safe random geometry generation module is used to predict the collision probability of path segments based on the perception road segment selector, the starting position of movement, the position of the moving target, the position of obstacles, and map environment information, and obtain a safe random geometry. The node heuristic value calculation module is used to calculate the node heuristic values ​​of the secure random geometry graph by using a global correction heuristic. The motion path planning module is used to plan the motion path based on node heuristics and a safe random geometry graph, through a differentiable A... The module performs path search to obtain the robot's motion planning path.

[0011] On the other hand, a robot motion planning device is provided, the robot motion planning device comprising: a processor; a memory storing computer-readable instructions, wherein when the computer-readable instructions are executed by the processor, any one of the robot motion planning methods inspired by the human brain's contextual semantic collaboration mechanism described above is implemented.

[0012] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement any of the above-described robot motion planning methods inspired by the human brain's contextual semantic coordination mechanism.

[0013] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention proposes a robot motion planning method inspired by the contextual semantic collaboration mechanism of the human brain. To improve the model's prediction accuracy of edge collision probabilities, a perceptual path selector is proposed, achieving efficient interaction and sharing between local edge features and global structural information, ensuring planning accuracy. Inspired by semantic memory providing global knowledge for contextual memory, a global correction heuristic is proposed. By introducing frequency domain features to provide full graph topological constraints, the model's prediction performance in high-dimensional weak graph structures is enhanced. Based on a dual-channel alignment learning method, the model's heuristic value prediction ability in high-dimensional weak graphs is improved, and the global path planning performance is enhanced. This invention is an efficient and flexible robot motion planning method that combines a graph neural network planner and a contextual semantic collaboration mechanism. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a robot motion planning method inspired by the human brain's contextual semantic collaboration mechanism, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of the framework of a sensing road segment selector provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the framework of a global correction heuristic provided in an embodiment of the present invention; Figure 4 This is a block diagram of a robot motion planning device inspired by the human brain's contextual semantic collaboration mechanism, provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the structure of a robot motion planning device provided in an embodiment of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0017] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0018] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0019] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] This invention provides a robot motion planning method inspired by the human brain's contextual semantic collaboration mechanism. This method can be implemented using a robot motion planning device, which can be a terminal or a server. Figure 1 The flowchart shown illustrates a robot motion planning method inspired by the human brain's contextual semantic coordination mechanism. The processing flow of this method may include the following steps: S1. Based on the robot intelligent operation platform, data is collected on the robot's motion tasks to obtain the starting position of the motion, the position of the target, the position of obstacles, and map environment information.

[0022] In one feasible implementation, in order to realize intelligent path planning for robots in complex industrial production, warehousing and logistics or service scenarios, and to reduce safety risks and operating costs in the actual testing process of robots, this invention builds a robot intelligent operation platform based on a simulation environment.

[0023] The robot intelligent operation platform can be used to learn models based on the current environment and to verify the effectiveness of the learned models in performing tasks. The platform consists of two parts: a simulation environment and a robot learning environment. The simulation environment simulates real-world manufacturing or service scenarios, constructing a virtual operating space that closely resembles actual applications. This simulation environment is built using PyBullet and can efficiently simulate the physical properties of the real environment, such as obstacle collisions. The simulation environment also includes an environment module for path planning tasks, a virtual robot module with different degrees of freedom and structural forms, and a virtual vision sensor module for real-time environmental information acquisition. Each path planning task includes obstacles, a starting vertex, and a target vertex.

[0024] The robot learning environment, serving as the intelligent control terminal of the aforementioned simulation environment, is installed on the digital control device of the actual production line or robot system. It is responsible for real-time analysis of input data and execution of corresponding intelligent learning algorithms. This robot learning environment sends decision-making instructions to the simulation environment to control the virtual robot in motion planning and path selection, and receives feedback information from the simulation environment after the robot interacts with the environment, dynamically optimizing subsequent decision-making processes.

[0025] S2. Based on the perceptual road segment selector, the collision probability of the path segment is predicted according to the starting position of the movement, the position of the moving target, the position of the obstacle, and the map environment information, so as to obtain a safe random geometry map.

[0026] Among them, the perceptual road segment selector uses the paths generated by Dijkstra's algorithm as training labels and is trained and optimized using the binary cross-entropy loss function; The sensing segment selector includes an encoder layer, a Transformer layer, a gated linear unit, and a selector output layer; The encoder layer includes node encoders, obstacle encoders, and edge encoders; The Transformer layer is a two-branch Transformer model structure; The selector output layer is a multilayer perceptron structure.

[0027] One feasible implementation method is, for example Figure 2 As shown, this invention relates to a perceptual road segment selector N based on a two-branch Transformer model. S The perceptual road segment selector N is trained using training data from the expert planner and labels. S Each planning problem in the training data { , It consists of obstacles and sampling configuration diagrams.

[0028] By uniformly sampling a series of nodes in the configuration space and constructing a random geometric graph using the K-Nearest Neighbors (K-NN) algorithm. ,in Through binary cross-entropy loss To train the algorithm, the confidence difference between the estimated values ​​and the true labels is minimized. Dijkstra's algorithm, acting as an expert planner, provides the true labels for the training problem. .

[0029] Optionally, based on the perceptual road segment selector, a path segment collision probability prediction is performed according to the starting position of the movement, the position of the moving target, the position of the obstacle, and map environment information to obtain a safe random geometry map, including: Based on the starting position of the motion, the position of the target motion, and the map environment information, the K-NN algorithm is used to construct the original random geometric graph; The original random geometry and obstacle positions are input into the perceptual road segment selector to predict the path segment collision probability and obtain the edge collision probability of the original random geometry. Based on a preset collision probability threshold, edge removal is performed on the original random geometry graph according to the edge collision probability to obtain a safe random geometry graph.

[0030] In one feasible implementation, a series of nodes are obtained through heuristic sampling in a collision-free configuration space. The original random geometric graph is constructed based on the K-Nearest Neighbors (K-NN) algorithm. Original random geometry RGG As input to the encoder, vertices and edges are encoded into the latent space. Each vertex... and each edge The embedding is represented as and ,in, and These are two different two-layer (Multilayer Perceptron, MLP) methods. In iterative obstacle coding... The obstacle encoding in the t-th iteration is represented by the following equation (1): (1); in, and These are the outputs of the vertices and edges after t-1 iterations of obstacle encoding, respectively. The representation layer is normalized, and O is the output of the obstacle after passing through the encoder. It is the MLP layer. These are the outputs of vertices and edges after t iterations of obstacle encoding, respectively. , for , Intermediate results output by the Transformer layer during the iterative update process.

[0031] Gated linear unit is represented as , Represents the learnable parameters. This is the bias vector.

[0032] The attention function for a node is expressed as follows (2): (2); in, These represent the key vector, query vector, and value vector for nodes and obstacles, respectively. Obstacle encoding for edges follows a similar process.

[0033] The obstacle encoder passes through T iterations. The information from the graph and obstacles is encoded into a joint embedding vector, represented as follows: .side The collision probability is obtained through a three-layer MLP with two-dimensional output. Calculation, expressed as It also removes edges with a collision probability higher than a pre-set threshold to obtain a safe random geometry. .

[0034] Optionally, the original random geometry and obstacle positions are input into a path segment selector to predict path segment collision probabilities, obtaining the edge collision probabilities of the original random geometry, including: Based on the original random geometry and obstacle locations, a latent space mapping is performed to obtain node embedding features, edge embedding features, and obstacle embedding features; Based on the attention mechanism, Transformer iterative encoding is performed according to node embedding features and obstacle embedding features to obtain node-obstacle features; Based on the attention mechanism, Transformer iterative encoding is performed according to edge embedding features and obstacle embedding features to obtain edge-obstacle features; The node-obstacle features and edge obstacle features are jointly encoded to obtain joint embedding features; Based on the joint embedding features, a multilayer perceptron is used to predict the collision probability and obtain the edge collision probability.

[0035] In one feasible implementation, the nodes and edges in the original random geometric graph, as well as the obstacles in the environment, are mapped to a unified latent space by an encoder to obtain node embedding features, edge embedding features, and obstacle embedding features.

[0036] The node embedding features and obstacle embedding features, as well as the edge embedding features and obstacle embedding features, are respectively input into the Transformer to obtain node features and edge features. At this point, the node and edge features contain obstacle information. The gated linear unit uses the output of the attention module in the Transformer to filter the gating information. After T iterations of the Transformer, the output is the node feature x^(T) and the edge feature y^(T) containing obstacle information.

[0037] Node features and edge features are encoded into a joint embedding vector u_{ij}, which can be viewed as the edge features of the joint node features. The final edge features u_{ij} are input into an MLP, and the output is a value in [0, 1], which is used by the model to predict the collision probability of edges in RGG.

[0038] S3. Based on the secure random geometry graph, calculate the heuristic values ​​of the nodes of the secure random geometry graph using a global correction heuristic.

[0039] Among them, the global correction heuristic is a graph neural network model based on the human brain's contextual semantic collaboration mechanism; The global correction heuristic is trained and optimized using a dual-channel alignment learning method; the dual-channel alignment learning method is used to strengthen the heuristic estimation of weak channels by connecting semantically similar nodes in the global scope. The global correction heuristic includes a spatial coding part, a spectral coding part, a graph neural network layer, and a heuristic output layer.

[0040] One feasible implementation method is, for example Figure 3 As shown, in order to improve the Global Alignment Heuristic (N... H Inspired by bidirectional interaction, this invention designs a dual-channel alignment learning (A2FL) method to train NH, which enhances the heuristic estimation capability of NH. A2FL connects nodes with shared semantic similarity from a global perspective, thereby facilitating heuristic estimation in weak channels. For NH and A... The module undergoes end-to-end joint training, which further improves the learning ability of heuristic values.

[0041] Training data is derived from graph search problem examples. Composition, in which, These represent the complete graph and the safe graph with collision edges removed, respectively. It is a binary vector containing the optimal path nodes generated by the expert planner.

[0042] The training loss for each channel can be calculated using A. The obtained path nodes and The L1 distance between them is represented. The loss functions for the two channels of the weak map and the full map are calculated as follows (3): (3); in, and They are respectively composed of differentiable A The module outputs path nodes based on both complete and weak graphs. The training loss penalizes overexplored nodes, guiding the search towards the optimal path.

[0043] To ensure semantic consistency between the two channels, this invention introduces an alignment loss, which... The node embeddings are mapped to the same latent space. Two MLP layers are used respectively. and To combine the node features of the two channels and Mapping to the underlying embedding and Alignment Loss (LA) is constructed using positive contrast difference (similarity of node features between two channels) and negative contrast difference (non-diagonal features within and between channels), as shown in Equation (4): (4); in , It is a cosine similarity function. The final objective function is... γ is given, and is used to adjust the magnitude of different losses.

[0044] Optionally, based on the secure random geometry graph, heuristic values ​​are calculated using a global correction heuristic to obtain the node heuristic values ​​of the secure random geometry graph, including: Based on the preset target nodes, spatial features are extracted from the secure random geometric graph to obtain spatial node features and spatial edge features. Based on the graph Fourier transform method, spectral features are extracted using the Laplace matrix based on the nodes of the secure random geometric graph to obtain the spectral domain node features; The spatial domain node features and the spectral domain node features are spliced ​​together to obtain the spliced ​​node features. Based on node splicing features and spatial edge features, a graph neural network is used to aggregate local information to obtain node aggregation features. Based on the node aggregation features, a multilayer perceptron is used to estimate the heuristic values ​​of the nodes.

[0045] One feasible implementation is based on a secure random set graph. and the given target node As input to this model, The nodes and edges in the array are also encoded into a latent space. Feature embedding The L2 distance between the node and the target node is additionally introduced, i.e. = ;side Feature embedding is defined as ,in and They are two different two-layer MLPs.

[0046] To enhance node feature representation, this invention uses Graph Fourier Transform (GFT) to extract spectral features. The Laplacian matrix is ​​defined as follows: ,in, It is the normalized adjacency matrix of the graph. Λ is the identity matrix, U is the matrix of eigenvectors, and Λ is the diagonal matrix of eigenvalues. Node features are represented by the Laplacian matrix. Recursive propagation. Therefore, nodes The spectral domain characteristics after k-order propagation are calculated as follows (5): (5); in, Let S denote the k-th order polynomial expansion of the Laplace matrix, and S denote the eigenspace generated by the principal components of the structure matrix, with learnable weights. and This allows for flexible reweighting of each feature subspace.

[0047] node spectral domain features With spatial characteristics Connecting to a concatenated vector can be represented as In Chinese, Graph Neural Networks (GNNs) work by aggregating local information from the neighboring nodes of each node. The node and edge embeddings are iteratively updated as follows: (6) (6); in, They are three different two-layer MLPs. This is the intermediate output of the node feature iterative update process. After L iterations, the node... The heuristic value is calculated as follows: ,in It is a three-layer MLP.

[0048] S4. Based on the node heuristic values ​​and the secure random geometry, through differentiable A... The module performs path search to obtain the robot's motion planning path.

[0049] Optionally, based on the node heuristic values ​​and a secure random geometry, through a differentiable A... The module performs path search to obtain the robot's motion planning path, including: Based on binary vector representation, open lists, closed lists, and cumulative costs are initialized according to node heuristics and a secure random geometry graph. Based on the open list and cumulative cost, the node is selected using a one-hot vector based on the node heuristic value. Move the selected node from the open list to the closed list, and update both the open and closed lists to obtain the updated open and closed lists. Based on the weighted adjacency matrix, the cumulative cost is calculated according to the safe random geometry and selected nodes to obtain the updated cumulative cost; Based on updating the closed list and updating the cumulative cost, the optimal path is searched according to the safe random geometry to obtain the robot's motion planning path.

[0050] In one feasible implementation, the heuristic value is used as A. The module's input, in The optimal path is then searched.

[0051] Introducing binary vectors and , respectively represent open lists. and closed lists In the process of selecting candidate nodes, one-hot vectors are used. This indicates the currently selected node. The candidate node with the lowest path cost is selected, and the process is as follows: (7) and (8): (7); (8); Where G is the cumulative cost, and H is a heuristic value estimated by a neural network. Represents element-wise product, used for labeling In the nodes, λ is a preset parameter. A function that returns a one-hot vector.

[0052] During the node expansion process, the matrix operation calculation of the candidate node's neighborhood is as follows (9): (9); in, Given a binary vector, mark the entry corresponding to an adjacent node as 1. It is a vector of all 1s. and Let f(x) represent the unweighted matrix and the weighted adjacency matrix of the edges, respectively.

[0053] Cumulative cost G and vectorized open list The update rule is as follows (10): (10); in, This represents the cumulative cost of the nodes associated with the path containing the selected node. This represents the distance from the selected node to each of its neighboring nodes. This represents the update of the cumulative cost of adjacent vertices.

[0054] In one feasible implementation, the present invention uses five key metrics to evaluate performance: the success rate of finding a collision-free path (SR), the number of collision checks (CK), the planning time (PT), the path cost (CO), and the penalized path cost (PO).

[0055] Here is the definition for each indicator: Success rate represents the percentage of 1000 test problems where a collision-free path is successfully found. Collision detection counts are performed by the collision detection module during the execution of the planning algorithm, checking whether each sampled node and the newly extended edge collide with an obstacle. Planning time is the time consumed in executing the planning algorithm. Path cost is the sum of the Euclidean path lengths of the collision-free paths found for each problem. Path cost with penalty terms is a larger penalty value assigned to failed paths to balance the fairness of path cost.

[0056] The performance comparison results of the proposed method and advanced methods on the above two datasets are shown in Table 1 (Comparison Table of Planning Performance on Maze Planning Dataset).

[0057] Stick3 is a 3-DOF (DoF) rod-shaped robot maze environment; Link8 is an 8-DOF (DoF) chain-shaped robot maze environment.

[0058] Table 1

[0059] This invention demonstrated significant performance across all tested simulation environments, achieving a success rate approaching 100%. Particularly in high-dimensional environments, it substantially reduced the number of collision checks and path costs. For methods with lower success rates, such as the Fast Exploration Random Tree Star algorithm and the Neural Exploration-Based Tree algorithm, which only successfully find paths in simpler example problems and thus have lower path costs, this invention introduces a penalty term for failed paths to balance the fairness of path cost comparisons. For this metric, this invention achieved optimal results in all simulation environments. Experimental results demonstrate that this invention exhibits robustness and stability for decision-making tasks under various complex environments.

[0060] This invention proposes a robot motion planning method inspired by the contextual semantic collaboration mechanism of the human brain. To improve the model's prediction accuracy of edge collision probabilities, a perceptual path selector is proposed, achieving efficient interaction and sharing between local edge features and global structural information, ensuring planning accuracy. Inspired by semantic memory providing global knowledge for contextual memory, a global correction heuristic is proposed. By introducing frequency domain features to provide full graph topological constraints, the model's prediction performance in high-dimensional weak graph structures is enhanced. Based on a dual-channel alignment learning method, the model's heuristic value prediction ability in high-dimensional weak graphs is improved, and the global path planning performance is enhanced. This invention is an efficient and flexible robot motion planning method that combines a graph neural network planner and a contextual semantic collaboration mechanism.

[0061] Figure 4 This is a block diagram illustrating a robot motion planning device inspired by the human brain's contextual semantic coordination mechanism, according to an exemplary embodiment. The device is used in a robot motion planning method inspired by the human brain's contextual semantic coordination mechanism. (Refer to...) Figure 4 The device includes an information acquisition module 410, a safe random geometry generation module 420, a node heuristic value calculation module 430, and a motion path planning module 440. Among them: The information acquisition module 410 is used to collect data on the robot's motion tasks based on the robot's intelligent operation platform, and obtain information such as the starting position of the motion, the position of the target motion, the position of obstacles, and the map environment. The safe random geometry generation module 420 is used to predict the collision probability of path segments based on the perceptual road segment selector, the starting position of the movement, the position of the moving target, the position of the obstacle, and the map environment information, and obtain a safe random geometry. The node heuristic value calculation module 430 is used to calculate the node heuristic values ​​of the secure random geometry graph by using a global correction heuristic. Motion path planning module 440 is used to plan motion paths based on node heuristics and a safe random geometry graph, via a differentiable A The module performs path search to obtain the robot's motion planning path.

[0062] Among them, the perceptual road segment selector uses the paths generated by Dijkstra's algorithm as training labels and is trained and optimized using the binary cross-entropy loss function; The sensing segment selector includes an encoder layer, a Transformer layer, a gated linear unit, and a selector output layer; The encoder layer includes node encoders, obstacle encoders, and edge encoders; The Transformer layer is a two-branch Transformer model structure; The selector output layer is a multilayer perceptron structure.

[0063] Optionally, the secure random geometry generation module 420 is further used for: Based on the starting position of the motion, the position of the target motion, and the map environment information, the K-NN algorithm is used to construct the original random geometric graph; The original random geometry and obstacle positions are input into the perceptual road segment selector to predict the path segment collision probability and obtain the edge collision probability of the original random geometry. Based on a preset collision probability threshold, edge removal is performed on the original random geometry graph according to the edge collision probability to obtain a safe random geometry graph.

[0064] Optionally, the secure random geometry generation module 420 is further used for: Based on the original random geometry and obstacle locations, a latent space mapping is performed to obtain node embedding features, edge embedding features, and obstacle embedding features; Based on the attention mechanism, Transformer iterative encoding is performed according to node embedding features and obstacle embedding features to obtain node-obstacle features; Based on the attention mechanism, Transformer iterative encoding is performed according to edge embedding features and obstacle embedding features to obtain edge-obstacle features; The node-obstacle features and edge obstacle features are jointly encoded to obtain joint embedding features; Based on the joint embedding features, a multilayer perceptron is used to predict the collision probability and obtain the edge collision probability.

[0065] Among them, the global correction heuristic is a graph neural network model based on the human brain's contextual semantic collaboration mechanism; The global correction heuristic is trained and optimized using a dual-channel alignment learning method; the dual-channel alignment learning method is used to strengthen the heuristic estimation of weak channels by connecting semantically similar nodes in the global scope. The global correction heuristic includes a spatial coding part, a spectral coding part, a graph neural network layer, and a heuristic output layer.

[0066] Optionally, the node heuristic value calculation module 430 is further used for: Based on the preset target nodes, spatial features are extracted from the secure random geometric graph to obtain spatial node features and spatial edge features. Based on the graph Fourier transform method, spectral features are extracted using the Laplace matrix based on the nodes of the secure random geometric graph to obtain the spectral domain node features; The spatial domain node features and the spectral domain node features are spliced ​​together to obtain the spliced ​​node features. Based on node splicing features and spatial edge features, a graph neural network is used to aggregate local information to obtain node aggregation features. Based on the node aggregation features, a multilayer perceptron is used to estimate the heuristic values ​​of the nodes.

[0067] Optionally, the motion path planning module 440 is further used for: Based on binary vector representation, open lists, closed lists, and cumulative costs are initialized according to node heuristics and a secure random geometry graph. Based on the open list and cumulative cost, the node is selected using a one-hot vector based on the node heuristic value. Move the selected node from the open list to the closed list, and update both the open and closed lists to obtain the updated open and closed lists. Based on the weighted adjacency matrix, the cumulative cost is calculated according to the safe random geometry and selected nodes to obtain the updated cumulative cost; Based on updating the closed list and updating the cumulative cost, the optimal path is searched according to the safe random geometry to obtain the robot's motion planning path.

[0068] This invention proposes a robot motion planning method inspired by the contextual semantic collaboration mechanism of the human brain. To improve the model's prediction accuracy of edge collision probabilities, a perceptual path selector is proposed, achieving efficient interaction and sharing between local edge features and global structural information, ensuring planning accuracy. Inspired by semantic memory providing global knowledge for contextual memory, a global correction heuristic is proposed. By introducing frequency domain features to provide full graph topological constraints, the model's prediction performance in high-dimensional weak graph structures is enhanced. Based on a dual-channel alignment learning method, the model's heuristic value prediction ability in high-dimensional weak graphs is improved, and the global path planning performance is enhanced. This invention is an efficient and flexible robot motion planning method that combines a graph neural network planner and a contextual semantic collaboration mechanism.

[0069] Figure 5 This is a structural schematic diagram of a robot motion planning device provided in an embodiment of the present invention, as shown below. Figure 5 As shown, the robot motion planning device may include the above-mentioned Figure 4 The illustrated robotic motion planning device is inspired by the contextual semantic coordination mechanism of the human brain. Optionally, the robotic motion planning device 510 may include a first processor 2001.

[0070] Optionally, the robot motion planning device 510 may also include a memory 2002 and a transceiver 2003.

[0071] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0072] The following is combined with Figure 5 A detailed introduction to each component of the robot motion planning device 510 is provided below: The first processor 2001 is the control center of the robot motion planning device 510. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0073] Optionally, the first processor 2001 can execute various functions of the robot motion planning device 510 by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0074] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 5 CPU0 and CPU1 are shown in the diagram.

[0075] In a specific implementation, as one example, the robot motion planning device 510 may also include multiple processors, for example... Figure 5 The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0076] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0077] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently, and may be accessed through the interface circuit of the robot motion planning device 510. Figure 5 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0078] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0079] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 5 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0080] Optionally, the transceiver 2003 can be integrated with the first processor 2001, or it can exist independently and be connected to the interface circuit of the robot motion planning device 510. Figure 5 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0081] It should be noted that, Figure 5 The structure of the robot motion planning device 510 shown does not constitute a limitation on the router. Actual knowledge structure recognition devices may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0082] Furthermore, the technical effects of the robot motion planning device 510 can be referenced from the technical effects of the robot motion planning method inspired by the human brain's contextual semantic collaboration mechanism described in the above method embodiments, and will not be repeated here.

[0083] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0084] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0085] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0086] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0087] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0088] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0089] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0090] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0091] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0092] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0093] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0094] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0095] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A robot motion planning method inspired by the contextual semantic collaboration mechanism of the human brain, characterized in that, The method includes: Based on the robot intelligent operation platform, data is collected on the robot's motion tasks to obtain the starting position of motion, the position of the target motion, the position of obstacles, and map environment information. Based on the perceptual road segment selector, the collision probability of the path segment is predicted according to the starting position of the movement, the position of the moving target, the position of the obstacle, and the map environment information, and a safe random geometry is obtained. The perceptual road segment selector uses paths generated by the Dijkstra algorithm as training labels and is trained and optimized using a binary cross-entropy loss function. The sensing segment selector includes an encoder layer, a Transformer layer, a gated linear unit, and a selector output layer; The encoder layer includes node encoders, obstacle encoders, and edge encoders; The Transformer layer is a two-branch Transformer model structure; The selector output layer is a multilayer perceptron structure; Based on the secure random geometry graph, the node heuristic values ​​of the secure random geometry graph are obtained by calculating the heuristic values ​​through a global correction heuristic. The global correction heuristic is based on a graph neural network model that utilizes the contextual semantic collaboration mechanism of the human brain. The global correction heuristic is trained and optimized using a dual-channel alignment learning method; the dual-channel alignment learning method is used to strengthen the heuristic estimation of weak channels by connecting semantically similar nodes in the global scope. The global correction heuristic includes a spatial coding part, a spectral coding part, a graph neural network layer, and a heuristic output layer; Based on node heuristics and a safe random geometry graph, a path search is performed using a differentiable A* module to obtain the robot's motion planning path.

2. The robot motion planning method inspired by the human brain's contextual semantic collaboration mechanism according to claim 1, characterized in that, The perceptual road segment selector predicts path segment collision probabilities based on the starting position of movement, the position of the moving target, the position of obstacles, and map environment information to obtain a safe random geometry map, including: Based on the starting position of the motion, the position of the target motion, and the map environment information, the K-NN algorithm is used to construct the original random geometric map. The original random geometry and obstacle locations are input into the perceptual road segment selector to predict the path segment collision probability, thereby obtaining the edge collision probability of the original random geometry. Based on a preset collision probability threshold, edge removal is performed on the original random geometry graph according to the edge collision probability to obtain a safe random geometry graph.

3. The robot motion planning method inspired by the human brain's contextual semantic collaboration mechanism according to claim 2, characterized in that, The step of inputting the original random geometry graph and obstacle positions into the perceptual road segment selector to predict the path segment collision probability and obtain the edge collision probability of the original random geometry graph includes: Based on the original random geometry and obstacle locations, a latent space mapping is performed to obtain node embedding features, edge embedding features, and obstacle embedding features; Based on the attention mechanism, Transformer iterative encoding is performed according to node embedding features and obstacle embedding features to obtain node-obstacle features; Based on the attention mechanism, Transformer iterative encoding is performed according to edge embedding features and obstacle embedding features to obtain edge-obstacle features; The node-obstacle features and edge obstacle features are jointly encoded to obtain joint embedding features; Based on the joint embedding features, a multilayer perceptron is used to predict the collision probability and obtain the edge collision probability.

4. The robot motion planning method inspired by the human brain's contextual semantic collaboration mechanism according to claim 1, characterized in that, The step of calculating heuristic values ​​for nodes of the secure random geometry graph using a global correction heuristic, based on the secure random geometry graph, includes: Based on the preset target nodes, spatial features are extracted from the secure random geometric graph to obtain spatial node features and spatial edge features. Based on the graph Fourier transform method, spectral features are extracted using the Laplace matrix based on the nodes of the secure random geometric graph to obtain the spectral domain node features; The spatial domain node features and the spectral domain node features are spliced ​​together to obtain the spliced ​​node features. Based on node splicing features and spatial edge features, a graph neural network is used to aggregate local information to obtain node aggregation features. Based on the node aggregation features, a multilayer perceptron is used to estimate the heuristic values ​​of the nodes.

5. The robot motion planning method inspired by the human brain's contextual semantic collaboration mechanism according to claim 1, characterized in that, The step of obtaining the robot motion planning path by performing path search through a differentiable A* module based on node heuristic values ​​and a secure random geometry graph includes: Based on binary vector representation, open lists, closed lists, and cumulative costs are initialized according to node heuristics and a secure random geometry graph. Based on the open list and cumulative cost, the node is selected using a one-hot vector based on the node heuristic value. Move the selected node from the open list to the closed list, and update both the open and closed lists to obtain the updated open and closed lists. Based on the weighted adjacency matrix, the cumulative cost is calculated according to the safe random geometry and selected nodes to obtain the updated cumulative cost; Based on updating the closed list and updating the cumulative cost, the optimal path is searched according to the safe random geometry to obtain the robot's motion planning path.

6. A robot motion planning device inspired by the human brain's contextual semantic collaboration mechanism, wherein the robot motion planning device is used to implement the robot motion planning method inspired by the human brain's contextual semantic collaboration mechanism as described in any one of claims 1-5, characterized in that, The device includes: The information acquisition module is used to collect data on the robot's motion tasks based on the robot's intelligent operation platform, and obtain information such as the starting position of the motion, the position of the target motion, the position of obstacles, and the map environment. The safe random geometry generation module is used to predict the collision probability of path segments based on the perception road segment selector, the starting position of movement, the position of the moving target, the position of obstacles, and map environment information, and obtain a safe random geometry. The node heuristic value calculation module is used to calculate the node heuristic values ​​of the secure random geometry graph by using a global correction heuristic. The motion path planning module is used to search for the robot's motion path through a differentiable A* module based on node heuristics and a safe random geometry.

7. A robot motion planning device, characterized in that, The robot motion planning device includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 5.

Images