Robot environment modeling and intelligent navigation method based on graph neural network

Abstract

The invention discloses a robot environment modeling and intelligent navigation method based on a graph neural network, and the method comprises the following steps: obtaining multi-source sensing data, and carrying out the time synchronization, thereby obtaining basic environment data; constructing an environment diagram with attributes based on the basic environment data; inputting the environment graph with attributes into a graph neural network coding module, and outputting a node embedding vector and a graph-level embedding vector; dividing the physical nodes into abstract nodes to form an abstract environment graph; calculating a multi-target cost weight for an abstract node by using the node embedding vector and the graph-level embedding vector; executing an improved Labeled RTDP algorithm on the abstract environment graph to obtain an optimal action of each abstract node; generating an abstract hierarchical navigation strategy according to the optimal action of each abstract node, and outputting robot linear velocity and angular velocity control quantity for executing navigation. According to the method, the environment modeling quality, the path planning efficiency and the real-time performance and safety of navigation decision making are effectively improved.

Description

Technical Field

[0001] This invention relates to the fields of robotics and artificial intelligence, and in particular to a method for robot environment modeling and intelligent navigation based on graph neural networks. Background Technology

[0002] With the widespread application of mobile robots in warehousing and logistics, service robots, intelligent security, and indoor inspection, achieving stable, safe, and efficient navigation in complex, dynamic, and unstructured environments has become a key technical challenge. Traditional environmental modeling methods often rely on grid maps, topological graphs, or manual modeling based on geometric features. These methods struggle to effectively represent multi-source information, semantic information, and structural relationships within the environment, resulting in insufficient environmental awareness for robots in complex scenarios and limiting the accuracy of navigation decisions. While some research has begun using convolutional neural networks to encode environmental data with the development of deep learning, convolutional structures still have significant limitations in handling multi-source information fusion and dynamic relationship modeling because they cannot directly process graph structure information.

[0003] In navigation planning, traditional search-based algorithms such as Dijkstra, A*, or D*, while computationally efficient, typically rely on fixed cost models, making it difficult to simultaneously address multi-objective optimization needs such as path length, safety, smoothness, and energy consumption. Furthermore, they are sensitive to environmental changes, requiring frequent replanning and incurring significant computational costs. To further improve planning capabilities, some studies have attempted to apply reinforcement learning or dynamic programming to navigation tasks, particularly Labeled RTDP algorithms, which have attracted attention due to their ability to converge quickly within a partial state space. However, the original Labeled RTDP assumes a static environment and a single cost; its solved labels cannot be revoked once set, lacking adaptability to dynamic environments. Its value updates depend on the complete state space, making it prone to excessive computational overhead in large-scale, complex environments.

[0004] Therefore, how to provide a robot environment modeling and intelligent navigation method based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a robot environment modeling and intelligent navigation method based on graph neural networks. This invention introduces graph neural networks to deeply encode the structure of the environment graph, and combines abstract environment graph construction, multi-objective cost modeling, and improved Labeled RTDP planning to achieve state space compression, local replanning in dynamic environments, and multi-objective adaptive optimization. This enables the generation of more stable, safe, smooth, and lower-energy-consumption navigation paths in complex scenarios. At the same time, it improves navigation decision efficiency through abstract hierarchical planning and physical layer path mapping, significantly enhancing the robot's autonomous navigation capability and overall robustness in dynamic and uncertain environments.

[0006] A robot environment modeling and intelligent navigation method based on graph neural networks according to an embodiment of the present invention includes the following steps: The robot acquires multi-source sensor data and performs time synchronization, noise filtering, and coordinate alignment to obtain basic environmental data. Based on basic environmental data, passable areas, static obstacles, dynamic obstacles, and semantic regions are constructed into an attributed environment graph with node attributes and edge attributes. The attributed environment graph is input into the graph neural network encoding module, which performs multi-round message passing and feature aggregation based on edge connection relationships, and outputs node embedding vectors and graph-level embedding vectors. Based on node embedding vectors and graph-level embedding vectors, connected and closely embedded physical nodes are divided into abstract nodes, and abstract passage costs and abstract risk values ​​are calculated to form an abstract environment graph. Calculate multi-objective cost weights for abstract nodes using node embedding vectors and graph-level embedding vectors; An improved Labeled RTDP algorithm is executed on an abstract environment graph. The stage cost is constructed using multi-objective cost weights for value updates, and solved labels are set and removed. When environmental changes are detected, local replanning is performed according to the graph distance radius to obtain the optimal action for each abstract node. An abstract hierarchical navigation strategy is generated based on the optimal actions of each abstract node, mapped to the physical environment graph to generate a local path, and the robot's linear velocity and angular velocity control values ​​are output for navigation execution.

[0007] Optionally, obtaining the basic environmental data specifically includes: The system simultaneously collects raw sensing data from lidar, camera, inertial measurement unit and odometer within a preset sampling period. The data is sorted according to the collection time. Interpolation and resampling methods are used to align the timestamps of different sensors to a unified time axis to form a multi-source sensing data sequence. Noise is filtered out from the multi-source sensing data sequence, and the filtered data are reorganized according to a unified time axis to obtain noise-suppressed multi-source data. The spatial points of the noise-suppressed multi-source data are sequentially transformed from the sensor coordinate system to the robot body coordinate system and then to the global coordinate system, and then combined with a unified time label and data source identifier to form basic environmental data.

[0008] Optionally, the construction of the attributed environment graph with node attributes and edge attributes specifically includes: Based on basic environmental data, spatial point data is classified in a global coordinate system, and spatial points belonging to passable areas, static obstacles, dynamic obstacles and semantic areas are respectively divided into discrete region elements. For each discrete region element, an environment graph node is generated, forming a node set, and node attributes are set for each node; Based on the node set, the connection relationship between nodes is established according to the comparison result of the spatial distance between any two nodes and the adjacency distance threshold. The nodes that meet the adjacency condition are used to establish the environment graph edge, and edge attributes are set for each edge. The node set and the edge set form an attributed environment graph with node attributes and edge attributes.

[0009] Optionally, the output process of the node embedding vector and the graph-level embedding vector specifically includes: The attributed environment graph is input into the graph neural network encoding module. For each node in the node set, the position attribute, size attribute, semantic category attribute, speed attribute and risk attribute are combined into a node feature vector according to a preset order. For each edge in the edge set, the edge length attribute and accessibility attribute are combined into an edge feature vector according to a preset order. All node feature vectors and edge feature vectors are used as input data for the graph neural network encoding module. In the graph neural network encoding module, message passing and feature aggregation based on edge connection relationships are performed on the attributed environment graph. For any target node, an adjacency message feature is generated according to the feature vectors of adjacent nodes and corresponding edges. The corresponding adjacency message feature is combined and updated with the node feature vector of the corresponding target node in the current round. Message passing and feature aggregation are repeated for a preset number of rounds until the node feature vectors of all nodes in the last round are obtained. The node feature vectors obtained in the last round are used as node embedding vectors. The reading operation is performed on all node embedding vectors within the node set in a summation manner to obtain graph-level embedding vectors used to characterize the overall features of the attributed environment graph.

[0010] Optionally, the formation of the abstract environment graph specifically includes: Based on the node embedding vector, all physical nodes in the attributed environment graph are traversed. For each physical node, a preset embedding feature extraction program is called to generate corresponding multi-scale embedding features. Physical nodes with edge connections are processed in pairs. Based on the multi-scale embedding features, a two-stage similarity determination process including candidate similarity calculation and context relevance verification is performed. When any pair of physical nodes simultaneously meets the connectivity requirements and the two-stage similarity threshold, the current physical node is assigned to the same physical node cluster, generating a physical node cluster set. For each physical node cluster, a corresponding abstract node is generated. All abstract nodes are combined into an abstract node set. Cross-cluster path relationship search is performed on any two abstract nodes in the abstract node set. When there is a cross-cluster connection path composed of physical edges in the corresponding physical node cluster, an abstract edge is established between the two abstract nodes. Joint feature fusion processing is performed on all physical edges involved in the cross-cluster path to generate abstract passage cost value and abstract risk value. The two are written into the abstract edge set as the passage cost attribute and risk attribute of the corresponding abstract edge, respectively. The abstract node set and the abstract edge set are organized into a unified structure to form an abstract environment graph.

[0011] Optionally, the calculation process of the multi-objective cost weights specifically includes: Read the corresponding node embedding vectors in the order of the abstract node numbers, concatenate the node embedding vector of each abstract node with the graph-level embedding vector at the same time in the feature dimension to generate the weight input feature vector of the abstract node, and write the weight input feature vectors of all abstract nodes into the weight calculation buffer in order. For each weight input feature vector in the weight calculation buffer, four scalar weights corresponding to path length cost, collision risk cost, motion smoothness cost and energy consumption cost are calculated in sequence. Normalization, upper and lower bound truncation and numerical stability correction are performed on the four scalar weights to generate multi-objective cost weights that correspond one-to-one with the abstract nodes. Optionally, obtaining the optimal action for each abstract node specifically includes: Based on the abstract environment graph and multi-objective cost weights, a corresponding value storage unit is established for each abstract node in the set of abstract nodes. The initial value of each abstract node is set to a preset initial value, and the solved labels of each abstract node are initialized to an unsolved state. Control parameters such as the maximum number of iterations, the value convergence residual threshold, the environment change threshold, and the graph distance radius are set to construct an improved Labeled RTDP algorithm solution environment for performing value updates and label management. In the improved Labeled RTDP solution environment, an abstract node access sequence is generated along the abstract edge starting from the current abstract node. For each abstract node in the access sequence, the basic cost information of the corresponding abstract edge is read sequentially according to the outgoing edge. The path length cost, collision risk cost, motion smoothness cost, and energy consumption cost on the abstract edge are calculated. Then, the multi-objective cost weight vector of the abstract node is read. The four basic costs and their corresponding weights are weighted and summed to obtain the stage cost value of the abstract edge. The stage cost value and the current value of the abstract node are combined according to the preset update rule to obtain the new node value. The difference between the new and old node values ​​is used as the value residual of the current abstract node. When the value residual of the same abstract node in multiple consecutive updates is less than the preset residual threshold and the environmental stability index of the region corresponding to the node is not lower than the stability threshold, the solved label of the abstract node is set to the solved state. The abstract node access, stage cost calculation, value update and label setting are repeated until the stopping condition is met. During the operation of the improved Labeled RTDP algorithm, environmental information is continuously monitored. When the change in the passage status, obstacle distribution, or risk attribute corresponding to any abstract edge in the abstract environment graph exceeds the environmental change threshold, the abstract node corresponding to the change position is taken as the center node. Based on the topology of the abstract environment graph, a set of abstract nodes whose distance from the center node graph is no greater than the preset graph distance radius is determined. The resolved labels of all abstract nodes in the set are revoked and the node values ​​are re-initialized according to preset rules. The value update and label setting process is re-executed within the local subgraph corresponding to the set of abstract nodes until the abstract nodes in the local subgraph meet the convergence condition. The current node value of each abstract node and the stage cost corresponding to each connected abstract edge are read. The stage costs of each candidate outgoing edge are compared, and the unique outgoing edge is selected as the optimal action for the abstract node according to the criterion of the minimum stage cost.

[0012] Optionally, the process of outputting the robot's linear and angular velocity control values ​​for navigation specifically includes: Determine the starting abstract node and target abstract node corresponding to the navigation starting point from the optimal action. Take the starting abstract node as the current node, select the unique outgoing edge according to the optimal action and transfer. Record the abstract nodes passed through in sequence, and use the recorded sequence of abstract nodes as the abstract level navigation trajectory. For each pair of adjacent abstract nodes in the abstract navigation trajectory, the corresponding physical node clusters are extracted, a physical node subgraph is constructed in the attributed environment graph, a local path search is performed based on the accessibility and risk attributes to obtain the corresponding local physical path, and all local physical paths are connected in sequence to generate a global physical path. The global physical path is discretized according to a preset path sampling interval, and corresponding linear velocity and angular velocity control commands are generated and sent to the robot chassis controller in sequence to drive the robot to perform navigation tasks along the global physical path.

[0013] The beneficial effects of this invention are: This invention achieves a unified expression of environmental semantics, structural relationships, and dynamic information by performing time synchronization, noise filtering, and coordinate alignment on multi-source sensing data, and further constructing an attributed environment graph with node and edge attributes.

[0014] This invention utilizes graph neural networks for multi-round message passing and feature aggregation, enabling the environment model to have stronger structural capture and semantic understanding capabilities, thus providing a high-quality environment representation for subsequent navigation planning.

[0015] This invention constructs an abstract environment graph based on node embedding vectors and graph-level embedding vectors, which effectively compresses the state space size, enabling the navigation algorithm to reduce computational burden and improve planning efficiency while maintaining the integrity of the environment structure.

[0016] This invention utilizes node embedding features and graph-level embedding features to adaptively generate multi-objective cost weights, enabling the navigation process to dynamically balance path length, safety, motion smoothness, and energy consumption. This breaks through the limitations of the fixed-cost model in traditional navigation algorithms and achieves adaptive adjustment at the multi-objective optimization level.

[0017] This invention introduces an improved Labeled RTDP algorithm, which integrates multi-objective cost weights and constructs stage costs, enabling value updates to fully consider environmental semantics and dynamic risk distribution. At the same time, by setting and revoking resolved labels, local replanning is performed within the graph distance radius when detecting environmental changes, effectively solving the defect of the original Labeled RTDP in adapting to dynamic environments and significantly improving the navigation strategy's agile response to environmental changes. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a flowchart of a robot environment modeling and intelligent navigation method based on graph neural networks proposed in this invention; Figure 2 This is a schematic diagram illustrating the multi-objective cost weight calculation of a robot environment modeling and intelligent navigation method based on graph neural networks proposed in this invention. Figure 3This is a schematic diagram of the algorithm structure of a robot environment modeling and intelligent navigation method based on graph neural networks proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1-3 A method for robot environment modeling and intelligent navigation based on graph neural networks includes the following steps: The robot acquires multi-source sensor data and performs time synchronization, noise filtering, and coordinate alignment to obtain basic environmental data. Based on basic environmental data, passable areas, static obstacles, dynamic obstacles, and semantic regions are constructed into an attributed environment graph with node attributes and edge attributes. The attributed environment graph is input into the graph neural network encoding module, which performs multi-round message passing and feature aggregation based on edge connection relationships, and outputs node embedding vectors and graph-level embedding vectors. Based on node embedding vectors and graph-level embedding vectors, connected and closely embedded physical nodes are divided into abstract nodes, and abstract passage costs and abstract risk values ​​are calculated to form an abstract environment graph. Multi-objective cost weights are calculated for abstract nodes using node embedding vectors and graph-level embedding vectors. These multi-objective cost weights include risk coefficients, path lengths, collision risks, motion smoothness, and energy consumption. An improved Labeled RTDP algorithm is executed on an abstract environment graph. The stage cost is constructed using multi-objective cost weights for value updates, and solved labels are set and removed. When environmental changes are detected, local replanning is performed according to the graph distance radius to obtain the optimal action for each abstract node. An abstract hierarchical navigation strategy is generated based on the optimal actions of each abstract node, mapped to the physical environment graph to generate a local path, and the robot's linear velocity and angular velocity control values ​​are output for navigation execution.

[0022] In this embodiment, obtaining the basic environmental data specifically includes: The system simultaneously collects raw sensing data from lidar, camera, inertial measurement unit and odometer within a preset sampling period. The data is sorted according to the collection time. Interpolation and resampling methods are used to align the timestamps of different sensors to a unified time axis to form a multi-source sensing data sequence. Noise is filtered out from the multi-source sensing data sequence, and the filtered data are reorganized according to a unified time axis to obtain noise-suppressed multi-source data. The spatial points of the noise-suppressed multi-source data are sequentially transformed from the sensor coordinate system to the robot body coordinate system and then to the global coordinate system, and then combined with a unified time label and data source identifier to form basic environmental data.

[0023] In this embodiment, the construction of the attributed environment graph with node attributes and edge attributes specifically includes: Based on basic environmental data, spatial point data is classified in a global coordinate system, and spatial points belonging to passable areas, static obstacles, dynamic obstacles and semantic areas are respectively divided into discrete region elements. An environment map node is generated for each discrete region element, forming a node set, and node attributes are set for each node. The node attributes include the three-dimensional position in the global coordinate system, region size, semantic category, velocity information corresponding to dynamic targets, and region risk information. Based on the node set, the connection relationship between nodes is established according to the comparison result of the spatial distance between any two nodes and the adjacency distance threshold. The nodes that meet the adjacency condition are used to establish the environment graph edge, and edge attributes are set for each edge. The node set and the edge set form an attributed environment graph with node attributes and edge attributes.

[0024] In this embodiment, the output process of the node embedding vector and the graph-level embedding vector specifically includes: The attributed environment graph is input into the graph neural network encoding module. For each node in the node set, the position attribute, size attribute, semantic category attribute, speed attribute, and risk attribute are combined into a node feature vector according to a preset order. For each edge in the edge set, the edge length attribute and traversability attribute are combined into an edge feature vector according to a preset order. All node feature vectors and edge feature vectors are used as input data for the graph neural network encoding module. The graph neural network encoding module is the environment perception and graph structure feature extraction subsystem in the robot computing system. In the graph neural network encoding module, message passing and feature aggregation based on edge connection relationships are performed on the attributed environment graph. For any target node, an adjacency message feature is generated according to the feature vectors of adjacent nodes and corresponding edges. The corresponding adjacency message feature is combined and updated with the node feature vector of the corresponding target node in the current round. Message passing and feature aggregation are repeated for a preset number of rounds until the node feature vectors of all nodes in the last round are obtained. The node feature vectors obtained in the last round are used as node embedding vectors. The reading operation is performed on all node embedding vectors within the node set in a summation manner to obtain graph-level embedding vectors used to characterize the overall features of the attributed environment graph.

[0025] In this embodiment, the formation of the abstract environment graph specifically includes: Based on the node embedding vector, all physical nodes in the attributed environment graph are traversed. For each physical node, a preset embedding feature extraction program is called to generate corresponding multi-scale embedding features. Physical nodes with edge connections are processed in pairs. Based on the multi-scale embedding features, a two-stage similarity determination process including candidate similarity calculation and context relevance verification is performed. When any pair of physical nodes simultaneously meets the connectivity requirements and the two-stage similarity threshold, the current physical node is assigned to the same physical node cluster, generating a physical node cluster set. For each physical node cluster, a corresponding abstract node is generated. All abstract nodes are combined into an abstract node set. Cross-cluster path relationship search is performed on any two abstract nodes in the abstract node set. When there is a cross-cluster connection path composed of physical edges in the corresponding physical node cluster, an abstract edge is established between the two abstract nodes. Joint feature fusion processing is performed on all physical edges involved in the cross-cluster path to generate abstract passage cost value and abstract risk value. The two are written into the abstract edge set as the passage cost attribute and risk attribute of the corresponding abstract edge, respectively. The abstract node set and the abstract edge set are organized into a unified structure. The unified structure organization process includes establishing a hierarchical mapping relationship between each abstract node and its corresponding physical node clusters, establishing an attribute mapping relationship between each abstract edge and its abstract travel cost attribute and abstract risk attribute, and performing topology consistency verification on the connection structure between abstract nodes to ensure that the reachability constraints of the original attributed environment graph are met, thus forming an abstract environment graph.

[0026] In this embodiment, the calculation process of the multi-objective cost weight specifically includes: Read the corresponding node embedding vectors in the order of the abstract node numbers, concatenate the node embedding vector of each abstract node with the graph-level embedding vector at the same time in the feature dimension to generate the weight input feature vector of the abstract node, and write the weight input feature vectors of all abstract nodes into the weight calculation buffer in order. For each weight input feature vector in the weight calculation buffer, four scalar weights corresponding to path length cost, collision risk cost, motion smoothness cost and energy consumption cost are calculated in sequence. Normalization, upper and lower bound truncation and numerical stability correction are performed on the four scalar weights to generate multi-objective cost weights that correspond one-to-one with the abstract nodes. In this embodiment, obtaining the optimal action for each abstract node specifically includes: Based on the abstract environment graph and multi-objective cost weights, a corresponding value storage unit is established for each abstract node in the set of abstract nodes. The initial value of each abstract node is set to a preset initial value, and the solved labels of each abstract node are initialized to an unsolved state. Control parameters such as the maximum number of iterations, the value convergence residual threshold, the environment change threshold, and the graph distance radius are set to construct an improved Labeled RTDP algorithm solution environment for performing value updates and label management. In the improved Labeled RTDP solution environment, an abstract node access sequence is generated starting from the current abstract node and proceeding along the abstract edges. For each abstract node in the access sequence, the basic cost information of the corresponding abstract edge is read sequentially according to the outgoing edges. The path length cost, collision risk cost, motion smoothness cost, and energy consumption cost on the abstract edge are calculated. Then, the multi-objective cost weight vector of the abstract node is read, and the four basic costs are weighted and summed with their corresponding weights to obtain the stage cost value of the abstract edge. The stage cost value is combined with the current value of the abstract node according to a preset update rule to obtain a new node value. The difference between the old and new node values ​​is used as the value residual of the current abstract node. When the value residual of the same abstract node in multiple consecutive updates is less than a preset residual threshold and the environmental stability index of the region corresponding to the node is not lower than the stability threshold, the solved label of the abstract node is set to the solved state. The abstract node access, stage cost calculation, value update, and label setting are repeated until the stopping condition is met. The environmental stability index is calculated based on the changes in the abstract passage cost, abstract risk value, and node embedding vector of the abstract node corresponding to the abstract edge within a preset time window. The stopping condition refers to the improved Labeled The number of RTDP iteration rounds has reached the preset maximum number of iterations; During the operation of the improved Labeled RTDP algorithm, environmental information is continuously monitored. When the change in the passage status, obstacle distribution, or risk attribute corresponding to any abstract edge in the abstract environment graph exceeds the environmental change threshold, the abstract node corresponding to the change position is taken as the center node. Based on the topology of the abstract environment graph, a set of abstract nodes whose graph distance from the center node is no greater than the preset graph distance radius is determined. The resolved labels of all abstract nodes in the set are revoked and the node values ​​are re-initialized according to preset rules. The value update and label setting process is re-executed within the local subgraph corresponding to the set of abstract nodes until the abstract nodes in the local subgraph meet the convergence condition. The current node value of each abstract node and the stage cost corresponding to each connected abstract edge are read. The stage costs of each candidate outgoing edge are compared, and the unique outgoing edge is selected as the optimal action for the abstract node according to the criterion of minimum stage cost. The local subgraph refers to the subgraph formed by the set of abstract nodes whose graph distance in the abstract environment graph is no greater than the preset graph distance radius and the internal connecting abstract edges, centered on the abstract node where the environment has changed.

[0027] This invention achieves efficient solutions to dynamic multi-objective navigation problems by introducing an improved Labeled RTDP algorithm onto an abstract environment graph. Compared to the limitation of traditional Labeled RTDP, which only handles single-objective costs, this invention combines four basic cost categories—path length, collision risk, motion smoothness, and energy consumption—with multi-objective cost weights generated by a graph neural network. This naturally integrates multi-objective optimization into the stage cost construction process, enabling value updates to simultaneously consider safety, efficiency, and smoothness, thereby significantly improving policy quality. Secondly, by introducing a label setting mechanism based on dual conditions of environmental stability index and value residual, the assignment of solved labels becomes more robust, avoiding misjudgments caused by slight environmental disturbances or value update fluctuations in the original algorithm. Furthermore, when detecting environmental changes, instead of performing global replanning, it limits the local influence range based on graph distance radius, only removing solved labels in local areas and reinitializing the corresponding node values. This significantly reduces the cost of resolving in dynamic environments, improves the algorithm's real-time response capability, and effectively enhances dynamic planning efficiency, multi-objective optimization capability, and environmental adaptability, providing a more reliable and efficient policy generation mechanism for intelligent navigation of robots in dynamic scenarios.

[0028] In this embodiment, the process of outputting the robot's linear velocity and angular velocity control quantities for navigation specifically includes: Determine the starting abstract node and target abstract node corresponding to the navigation starting point from the optimal action. Take the starting abstract node as the current node, select the unique outgoing edge according to the optimal action and transfer. Record the abstract nodes passed through in sequence, and use the recorded sequence of abstract nodes as the abstract level navigation trajectory. For each pair of adjacent abstract nodes in the abstract navigation trajectory, the corresponding physical node clusters are extracted, a physical node subgraph is constructed in the attributed environment graph, a local path search is performed based on the accessibility and risk attributes to obtain the corresponding local physical path, and all local physical paths are connected in sequence to generate a global physical path. The global physical path is discretized according to a preset path sampling interval, and corresponding linear velocity and angular velocity control commands are generated and sent to the robot chassis controller in sequence to drive the robot to perform navigation tasks along the global physical path.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a typical indoor mobile robot autonomous inspection task. The environment was a multi-corridor connected area approximately 150 meters long, including open walkways, glass-partitioned office areas, unattended equipment rooms, and densely populated dynamic passageways. This scenario is characterized by frequent dynamic obstacles, complex environmental semantics, and rapid changes in local area structure. Traditional A* and D* Lite planning methods in such scenarios are prone to unstable path planning, frequent replanning, and uneven navigation trajectories, failing to meet the comprehensive requirements of real-time performance and safety in inspection tasks. This invention constructs an attributed environment graph and utilizes graph neural networks to extract environmental structural features, enabling the robot to simultaneously understand spatial topological relationships and semantic region information, thereby forming an abstract representation that more closely approximates the real-world environmental structure during actual inspections.

[0030] In practical applications, the robot first collects environmental data through multi-source sensors, including laser point clouds, depth images, visual semantic segmentation results, and its own odometry data. After time synchronization and noise filtering, these data forms basic environmental data. This invention then constructs an attributed environmental graph with node and edge attributes, which fully covers the passable areas and obstacle distribution of the inspection zone. Subsequently, this invention inputs the environmental graph into a graph neural network encoding module and performs five rounds of message passing and feature aggregation to generate node embedding vectors and graph-level embedding vectors. This enables the robot to extract hierarchical features from the environment's geometric structure, semantic categories, and dynamic changes. Based on this embedding information, physical nodes are further clustered into abstract nodes. An abstract environmental graph is formed by calculating abstract passage costs and abstract risk values. This allows the planning algorithm to significantly reduce the number of states while maintaining environmental structural consistency, thereby improving planning efficiency.

[0031] During the inspection process, this invention generates adaptive weight combinations of path length, safety, smoothness, and energy consumption for abstract nodes through multi-objective cost weights. This allows the navigation strategy to automatically adjust its optimization direction based on environmental semantics and dynamic risks. For example, it increases risk weights in densely populated areas, improves smoothness weights in narrow corridors, and enhances energy consumption weights in long-distance travel areas, thereby achieving multi-objective navigation optimization. An improved Labeled RTDP algorithm is run on the abstract environment graph. By establishing node value storage units and combining them with stage costs for value updates, the robot can quickly obtain the optimal action for each abstract node. Since people frequently pass through the inspection area, and the travel environment is random, this invention automatically triggers local replanning when the deviation between the actual risk and the abstract risk in a local area exceeds a threshold. Value updates are only re-executed for twenty to thirty abstract nodes whose distance from the changed area graph does not exceed a set radius, thus avoiding the time-consuming global replanning process required by traditional planning methods.

[0032] To verify the effectiveness of this invention, a comparative experiment will be conducted with the traditional A* and D* Lite algorithms in the same scenario. The specific experimental data is shown in Table 1. Table 1. Performance Comparison of the Invention Method and Traditional Methods in Inspection Scenarios

[0033] Table 1 shows that the overall performance of this invention in navigation tasks is significantly superior to traditional methods. The path generated by this invention is more than 10m shorter than that of traditional algorithms, the total planning time is reduced by nearly 200ms, and the time for a single replanning operation is reduced to less than one-third of that of traditional methods, resulting in a significant improvement in the system's response speed in dynamic environments. Due to the adoption of an abstract environment graph and a local replanning mechanism, the number of replanning operations is reduced by approximately 60%, and the number of emergency stops is reduced by nearly 85%, demonstrating a more stable and safer navigation process. Furthermore, the energy consumption index of this invention is reduced by approximately 0.1 compared to traditional methods, and the smoothness score is improved by more than 0.15, indicating that while ensuring safety, trajectory quality and operational efficiency are also comprehensively enhanced. These data fully demonstrate that this invention achieves significant improvements in path quality, efficiency, stability, and energy consumption control.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for robot environment modeling and intelligent navigation based on graph neural networks, characterized in that, Includes the following steps: The robot acquires multi-source sensor data and performs time synchronization, noise filtering, and coordinate alignment to obtain basic environmental data. Based on basic environmental data, passable areas, static obstacles, dynamic obstacles, and semantic regions are constructed into an attributed environment graph with node attributes and edge attributes. The attributed environment graph is input into the graph neural network encoding module, which performs multi-round message passing and feature aggregation based on edge connection relationships, and outputs node embedding vectors and graph-level embedding vectors. Based on node embedding vectors and graph-level embedding vectors, connected and closely embedded physical nodes are divided into abstract nodes, and abstract passage costs and abstract risk values ​​are calculated to form an abstract environment graph. Calculate multi-objective cost weights for abstract nodes using node embedding vectors and graph-level embedding vectors; An improved Labeled RTDP algorithm is executed on an abstract environment graph. The stage cost is constructed using multi-objective cost weights for value updates, and solved labels are set and removed. When environmental changes are detected, local replanning is performed according to the graph distance radius to obtain the optimal action for each abstract node. An abstract hierarchical navigation strategy is generated based on the optimal actions of each abstract node, mapped to the physical environment graph to generate a local path, and the robot's linear velocity and angular velocity control values ​​are output for navigation execution.

2. The robot environment modeling and intelligent navigation method based on graph neural networks according to claim 1, characterized in that, The acquisition of the basic environmental data specifically includes: The system simultaneously collects raw sensing data from lidar, camera, inertial measurement unit and odometer within a preset sampling period. The data is sorted according to the collection time. Interpolation and resampling methods are used to align the timestamps of different sensors to a unified time axis to form a multi-source sensing data sequence. Noise is filtered out from the multi-source sensing data sequence, and the filtered data are reorganized according to a unified time axis to obtain noise-suppressed multi-source data. The spatial points of the noise-suppressed multi-source data are sequentially transformed from the sensor coordinate system to the robot body coordinate system and then to the global coordinate system, and then combined with a unified time label and data source identifier to form basic environmental data.

3. The robot environment modeling and intelligent navigation method based on graph neural networks according to claim 1, characterized in that, The construction of the attributed environment graph with node and edge attributes specifically includes: Based on basic environmental data, spatial point data is classified in a global coordinate system, and spatial points belonging to passable areas, static obstacles, dynamic obstacles and semantic areas are respectively divided into discrete region elements. For each discrete region element, an environment graph node is generated, forming a node set, and node attributes are set for each node; Based on the node set, the connection relationship between nodes is established according to the comparison result of the spatial distance between any two nodes and the adjacency distance threshold. The nodes that meet the adjacency condition are used to establish the environment graph edge, and edge attributes are set for each edge. The node set and the edge set form an attributed environment graph with node attributes and edge attributes.

4. The robot environment modeling and intelligent navigation method based on graph neural networks according to claim 1, characterized in that, The output process of the node embedding vector and the graph-level embedding vector specifically includes: The attributed environment graph is input into the graph neural network encoding module. For each node in the node set, the position attribute, size attribute, semantic category attribute, speed attribute and risk attribute are combined into a node feature vector according to a preset order. For each edge in the edge set, the edge length attribute and accessibility attribute are combined into an edge feature vector according to a preset order. All node feature vectors and edge feature vectors are used as input data for the graph neural network encoding module. In the graph neural network encoding module, message passing and feature aggregation based on edge connection relationships are performed on the attributed environment graph. For any target node, an adjacency message feature is generated according to the feature vectors of adjacent nodes and corresponding edges. The corresponding adjacency message feature is combined and updated with the node feature vector of the corresponding target node in the current round. Message passing and feature aggregation are repeated for a preset number of rounds until the node feature vectors of all nodes in the last round are obtained. The node feature vectors obtained in the last round are used as node embedding vectors. The reading operation is performed on all node embedding vectors within the node set in a summation manner to obtain graph-level embedding vectors used to characterize the overall features of the attributed environment graph.

5. The robot environment modeling and intelligent navigation method based on graph neural networks according to claim 1, characterized in that, The formation of the abstract environment diagram specifically includes: Based on the node embedding vector, all physical nodes in the attributed environment graph are traversed. For each physical node, a preset embedding feature extraction program is called to generate corresponding multi-scale embedding features. Physical nodes with edge connections are processed in pairs. Based on the multi-scale embedding features, a two-stage similarity determination process including candidate similarity calculation and context relevance verification is performed. When any pair of physical nodes simultaneously meets the connectivity requirements and the two-stage similarity threshold, the current physical node is assigned to the same physical node cluster, generating a physical node cluster set. For each physical node cluster, a corresponding abstract node is generated. All abstract nodes are combined into an abstract node set. Cross-cluster path relationship search is performed on any two abstract nodes in the abstract node set. When there is a cross-cluster connection path composed of physical edges in the corresponding physical node cluster, an abstract edge is established between the two abstract nodes. Joint feature fusion processing is performed on all physical edges involved in the cross-cluster path to generate abstract passage cost value and abstract risk value. The two are written into the abstract edge set as the passage cost attribute and risk attribute of the corresponding abstract edge, respectively. The abstract node set and the abstract edge set are organized into a unified structure to form an abstract environment graph.

6. The robot environment modeling and intelligent navigation method based on graph neural networks according to claim 1, characterized in that, The calculation process of the multi-objective cost weight specifically includes: Read the corresponding node embedding vectors in the order of the abstract node numbers, concatenate the node embedding vector of each abstract node with the graph-level embedding vector at the same time in the feature dimension to generate the weight input feature vector of the abstract node, and write the weight input feature vectors of all abstract nodes into the weight calculation buffer in order. For each weight input feature vector in the weight calculation buffer, four scalar weights corresponding to path length cost, collision risk cost, motion smoothness cost, and energy consumption cost are calculated in sequence. Normalization, upper and lower bound truncation, and numerical stability correction are performed on the four scalar weights to generate multi-objective cost weights that correspond one-to-one with the abstract nodes.

7. The robot environment modeling and intelligent navigation method based on graph neural networks according to claim 1, characterized in that, The optimal action for each abstract node is obtained specifically through: Based on the abstract environment graph and multi-objective cost weights, a corresponding value storage unit is established for each abstract node in the set of abstract nodes. The initial value of each abstract node is set to a preset initial value, and the solved labels of each abstract node are initialized to an unsolved state. Control parameters such as the maximum number of iterations, the value convergence residual threshold, the environment change threshold, and the graph distance radius are set to construct an improved Labeled RTDP algorithm solution environment for performing value updates and label management. In the improved Labeled RTDP solution environment, an abstract node access sequence is generated along the abstract edge starting from the current abstract node. For each abstract node in the access sequence, the basic cost information of the corresponding abstract edge is read sequentially according to the outgoing edge. The path length cost, collision risk cost, motion smoothness cost, and energy consumption cost on the abstract edge are calculated. Then, the multi-objective cost weight vector of the abstract node is read. The four basic costs and their corresponding weights are weighted and summed to obtain the stage cost value of the abstract edge. The stage cost value and the current value of the abstract node are combined according to the preset update rule to obtain the new node value. The difference between the new and old node values ​​is used as the value residual of the current abstract node. When the value residual of the same abstract node in multiple consecutive updates is less than the preset residual threshold and the environmental stability index of the region corresponding to the node is not lower than the stability threshold, the solved label of the abstract node is set to the solved state. The abstract node access, stage cost calculation, value update and label setting are repeated until the stopping condition is met. During the operation of the improved Labeled RTDP algorithm, environmental information is continuously monitored. When the change in the passage status, obstacle distribution, or risk attribute corresponding to any abstract edge in the abstract environment graph exceeds the environmental change threshold, the abstract node corresponding to the change position is taken as the center node. Based on the topology of the abstract environment graph, a set of abstract nodes whose distance from the center node graph is no greater than the preset graph distance radius is determined. The resolved labels of all abstract nodes in the set are revoked and the node values ​​are re-initialized according to preset rules. The value update and label setting process is re-executed within the local subgraph corresponding to the set of abstract nodes until the abstract nodes in the local subgraph meet the convergence condition. The current node value of each abstract node and the stage cost corresponding to each connected abstract edge are read. The stage costs of each candidate outgoing edge are compared, and the unique outgoing edge is selected as the optimal action for the abstract node according to the criterion of the minimum stage cost.

8. The robot environment modeling and intelligent navigation method based on graph neural networks according to claim 1, characterized in that, The process of using the output robot linear velocity and angular velocity control quantities to perform navigation specifically includes: Determine the starting abstract node and target abstract node corresponding to the navigation starting point from the optimal action. Take the starting abstract node as the current node, select the unique outgoing edge according to the optimal action and transfer. Record the abstract nodes passed through in sequence, and use the recorded sequence of abstract nodes as the abstract level navigation trajectory. For each pair of adjacent abstract nodes in the abstract navigation trajectory, the corresponding physical node clusters are extracted, a physical node subgraph is constructed in the attributed environment graph, a local path search is performed based on the accessibility and risk attributes to obtain the corresponding local physical path, and all local physical paths are connected in sequence to generate a global physical path. The global physical path is discretized according to a preset path sampling interval, and corresponding linear velocity and angular velocity control commands are generated and sent to the robot chassis controller in sequence to drive the robot to perform navigation tasks along the global physical path.

Images