Robot optimal path planning method based on 3D environment semantics and dynamic obstacle avoidance

Abstract

This disclosure provides a robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance, including: acquiring workspace perception data, and performing 3D mesh division and semantic classification on the workspace perception data to construct a semantic partitioned accessibility model. Based on the semantic partitioned accessibility model, spatial overlay analysis is performed on the obstacle distribution within the workspace to identify closed regions and generate a semantic closed region distribution map. The robot's initial path is segmented according to the semantic closed region distribution map, and semantic access windows are constructed within the closed regions to generate a candidate path sequence constrained by the semantic access windows. The candidate path sequences are sorted using a semantic priority factor to obtain the semantically guided optimal path sequence. When a path segment in the optimal path sequence enters a closed region, the path segment is delayed or reconstructed using the semantic access window to output a path planning result that satisfies the semantic access window constraints.

Description

Technical Field

[0001] This disclosure belongs to the field of artificial intelligence technology, and more specifically, relates to a robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance. Background Technology

[0002] In automated assembly lines or multi-robot collaborative handling scenarios, a large number of moving objects can easily form local dynamic closed areas. The robot path cost field is prone to continuous high occupancy areas in a short period of time, causing the path search results to frequently revert or fall into local stagnation, which makes it difficult for existing robot handling technologies to form executable paths. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to resolve the aforementioned deficiencies and propose a robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance.

[0004] The present invention adopts the following technical solution.

[0005] The first aspect of this invention discloses a robot optimal path planning method based on 3D environment semantics and dynamic obstacle avoidance, the method comprising:

[0006] Acquire workspace perception data, and perform three-dimensional mesh division and semantic classification on the workspace perception data to construct a semantic partition accessibility model;

[0007] Based on the semantic partitioning accessibility model, spatial overlay analysis is performed on the distribution of obstacles in the work space to identify closed areas and generate a semantic closed area distribution map.

[0008] The robot's initial path is segmented according to the semantic closed region distribution map, and a semantic access window is constructed within the closed region to generate a candidate path sequence constrained by the semantic access window.

[0009] The candidate path sequences are sorted by a semantic priority factor to obtain the optimal semantically guided path sequence.

[0010] When a path segment in the optimal path sequence enters a closed region, the path segment is delayed or reconstructed through the semantic access window to output a path planning result that satisfies the semantic access window constraints.

[0011] Furthermore, the step of acquiring workspace perception data and performing three-dimensional mesh partitioning and semantic classification on the workspace perception data to construct a semantic partition accessibility model includes:

[0012] Visual images and depth ranging data of the workspace are acquired through image acquisition equipment and a ranging module, and distortion correction and alignment calibration are performed on the visual images and depth ranging data.

[0013] The work space is divided into three-dimensional meshes according to the preset voxel side lengths, and the object category, region function and behavior state corresponding to the visual image are identified by the visual model so as to map the object category, region function and behavior state to the corresponding spatial units respectively.

[0014] The spatial unit is obtained by dividing the work space into three-dimensional meshes, the visual model is a cascaded structure of a target detection model and a semantic segmentation model, and the work space perception data includes the visual image and depth ranging data.

[0015] Furthermore, the step of acquiring workspace perception data and performing three-dimensional mesh division and semantic classification on the workspace perception data to construct a semantic partitioning accessibility model further includes:

[0016] The semantic scores of different semantic categories in each spatial unit are statistically analyzed. The semantic category with the highest semantic score is selected as the semantic label of the corresponding spatial unit. The spatial units are then aggregated based on the adjacency relationship between different spatial units and the similarity between semantic labels and behavioral states to obtain multiple semantic partitions.

[0017] Calculate the static access capability and dynamic access capability corresponding to each semantic partition, and merge the static access capability and dynamic access capability into the corresponding semantic partition to obtain the semantic partition access capability model.

[0018] Furthermore, based on the semantic partitioning accessibility model, the spatial overlay analysis of obstacle distribution within the work space is performed to identify enclosed areas and generate a semantic enclosed area distribution map, including:

[0019] The moving objects in the work space corresponding to each semantic partition are continuously identified and tracked to obtain the number, location and behavior status of the moving objects, and the number, location and behavior status of the moving objects are bound to the dynamic accessibility to obtain a dynamic obstacle semantic coupling sequence.

[0020] Temporal statistics are performed on the dynamic obstacle semantic coupling sequence to extract the instantaneous occupancy features of each semantic partition within a set time window, and the closure index of the corresponding semantic partition is calculated based on the instantaneous occupancy features.

[0021] The semantic partitions are classified according to the closure index to generate the semantic closure region distribution map.

[0022] Furthermore, the step of segmenting the robot's initial path according to the semantic closed region distribution map and constructing a semantic passage window within the closed region to generate a candidate path sequence constrained by the semantic passage window includes:

[0023] The robot's initial path is mapped segment by segment to the semantically closed region distribution map, and each path segment is divided into a passable segment and a semantically restricted segment;

[0024] The semantic access window corresponding to the semantically restricted segment is determined based on the dynamic access capability and closure degree index, and multiple candidate access paths are constructed based on the semantic access window;

[0025] The candidate paths and segments are spliced ​​together according to the passage order of each segment in the robot's initial path, and the spliced ​​paths are sorted and trimmed to obtain the candidate path sequence.

[0026] Furthermore, the step of sorting the candidate path sequences using a semantic priority factor to obtain the semantically guided optimal path sequence includes:

[0027] Semantic priority expansion is performed on the semantic partitions where each path segment in the candidate path sequence is located to construct a path segment evaluation table corresponding to the candidate path sequence, and the semantic priority factor is constructed based on the path segment evaluation table;

[0028] The candidate path sequences are sorted according to the semantic priority factor to obtain the semantic priority ranking result of the candidate path sequences, and the candidate path sequence with the highest ranking is selected from the semantic priority ranking result of the candidate path sequences as the initial optimal path sequence.

[0029] Furthermore, the step of sorting the candidate path sequences using a semantic priority factor to obtain the semantically guided optimal path sequence also includes:

[0030] When executing the initial optimal path sequence, the updated semantic closed region distribution map is received in real time, and the path segment to be executed is reviewed to determine whether the path segment to be executed meets the constraints of the semantic passage window.

[0031] When the path segment to be executed does not meet the constraints of the semantic passage window, the path segment with the highest semantic priority is selected from the candidate path sequence to replace the path segment to be executed, thereby obtaining the optimal path sequence guided by semantics.

[0032] Furthermore, when a path segment in the optimal path sequence enters a closed region, the path segment is delayed or reconstructed through the semantic access window to output a path planning result that satisfies the semantic access window constraints, including:

[0033] The execution order of each path segment in the optimal path sequence is continuously checked to obtain the continuous check results. The continuous check results are then used for risk screening to identify risky path segments that enter the closed area.

[0034] Determine whether the risk path segment has a semantic access window, and if the risk path segment has the semantic access window, perform delayed execution processing on the risk path segment; otherwise, replace the risk path segment.

[0035] The path planning results include risky path segments that have undergone delayed execution processing or replacement.

[0036] The second aspect of this invention discloses a robot optimal path planning device based on 3D environmental semantics and dynamic obstacle avoidance, used to implement the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance as described in any of the first aspects, the device comprising:

[0037] The access model construction module is used to acquire workspace perception data and perform three-dimensional mesh division and semantic classification on the workspace perception data to construct a semantic partition access capability model.

[0038] The closed area identification module is used to perform spatial overlay analysis on the distribution of obstacles in the work space based on the semantic partition accessibility model, so as to identify closed areas and generate a semantic closed area distribution map.

[0039] The candidate path generation module is used to segment the robot's initial path according to the semantic closed region distribution map, and construct a semantic access window within the closed region to generate a candidate path sequence constrained by the semantic access window;

[0040] The optimal path filtering module is used to sort the candidate path sequence by a semantic priority factor in order to obtain the semantically guided optimal path sequence.

[0041] The path planning output module is used to delay or reconstruct the path segment through the semantic access window when the path segment in the optimal path sequence enters the closed area, so as to output the path planning result that satisfies the constraints of the semantic access window.

[0042] A third aspect of the present invention discloses a terminal, including a processor and a storage medium;

[0043] The storage medium is used to store instructions;

[0044] The processor is configured to operate according to the instructions to perform the steps of the method described in the first aspect.

[0045] A fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0046] The beneficial effects of the present invention are as follows: Compared with the prior art, the present invention has the following advantages:

[0047] (1) This invention divides the work space into three-dimensional grids and performs semantic classification of spatial units based on work functions and behavioral characteristics to obtain semantic partitions. At the same time, each semantic partition is assigned a capability value that can form passage gaps per unit time. Then, spatial overlay analysis is performed on the distribution of dynamic obstacles to identify whether local areas form semantically temporary closed areas or high-frequency passage areas. Then, the degree of closure index within the semantic partition is calculated, and a semantic closed area distribution map is output. This can transform the dynamic behavior information obtained by industrial vision into a quantitative basis for temporary closed areas or high-frequency passage areas, which essentially avoids the phenomenon of frequent regression or local stagnation in robot path search results.

[0048] (2) This invention segments the original path using a semantic closed region distribution map. Within the closed region, obstacle avoidance is no longer performed directly. Instead, temporal and spatial passage windows are generated based on semantic traversability, and a candidate path set is constructed, outputting a set of candidate path sequences under semantic constraints. Simultaneously, a semantic priority factor is introduced to sort the paths, prioritizing path segments with high traversability and low closure. During execution, paths are locally replaced and spliced ​​based on real-time semantic partition changes, outputting a semantically guided dynamic optimal path sequence. Finally, path continuity detection is performed. When a path segment is detected to enter a highly closed semantic region, a semantic passage window mechanism is triggered for delayed passage or local replanning, generating an executable path result without path locking. This invention addresses the technical problem of path locking caused by dynamic obstacles forming locally closed regions in high-density collaborative operation scenarios, achieving the generation of globally optimal executable paths even under conditions of dense interference from multiple moving bodies. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention.

[0050] Figure 2 This is a schematic diagram of the robot optimal path planning device based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention. Detailed Implementation

[0051] The present application will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and should not be construed as limiting the scope of protection of the present application.

[0052] Considering the dynamic obstacle avoidance path locking problem in high-density collaborative operation scenarios, such as automated assembly lines or multi-robot collaborative handling scenarios, a large number of moving bodies form locally dynamic closed areas. The path cost field experiences continuous high occupancy areas within a short period, causing path search results to frequently revert or become stuck in local stagnation, preventing the formation of an executable path. Based on this, such as... Figure 1 As shown, this invention discloses a robot optimal path planning method based on 3D environment semantics and dynamic obstacle avoidance, including the following steps:

[0053] Step S110: Obtain the work space perception data, and perform three-dimensional grid division and semantic classification on the work space perception data to construct a semantic partition access capability model.

[0054] In some embodiments, the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention includes the following steps in step S110:

[0055] Step S111: Obtain visual images and depth ranging data of the work space through image acquisition equipment and ranging module, and perform distortion correction and alignment calibration on the visual images and depth ranging data.

[0056] Step S112: Divide the work space into a three-dimensional mesh according to the preset voxel side length, and identify the object category, region function and behavior state corresponding to the visual image through the visual model, so as to map the object category, region function and behavior state to the corresponding spatial unit respectively.

[0057] Among them, the spatial unit is obtained by dividing the work space into three-dimensional grids, the visual model is a series structure of target detection model and semantic segmentation model, and the work space perception data includes visual images and depth ranging data.

[0058] In some embodiments, the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention further includes the following steps in step S110:

[0059] Step S113: Calculate the semantic scores of different semantic categories in each spatial unit, select the semantic category with the highest semantic score as the semantic label of the corresponding spatial unit, and aggregate the spatial units according to the adjacency relationship of different spatial units and the similarity between semantic labels and behavioral states to obtain multiple semantic partitions.

[0060] In some embodiments, the process of aggregating spatial units based on their adjacency relationships, semantic tags, and behavioral states is as follows: First, adjacent spatial units are selected. Then, it is determined whether their main semantic tags are consistent, and a difference constraint is applied to the behavioral activity. If the consistency and threshold conditions are met, they are merged into the same semantic partition. For example, multiple consecutive main channel spatial units with continuous traversal behavior are aggregated into one main channel semantic partition.

[0061] Step S114: Calculate the static access capability and dynamic access capability corresponding to each semantic partition, and merge the static access capability and dynamic access capability into the corresponding semantic partition to obtain the semantic partition access capability model.

[0062] In a specific embodiment, the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention includes steps 1 to 5:

[0063] Step 1: Perform 3D environment semantic partitioning and access attribute modeling.

[0064] The workspace is divided into a three-dimensional mesh, and the spatial units are semantically classified based on work functions and behavioral characteristics to obtain a set of semantic partitions. Simultaneously, each semantic partition is assigned access attributes, including static access capacity and dynamic access capacity. Dynamic access capacity is defined as the ability to form passage gaps per unit time, including the following sub-steps:

[0065] Sub-step 1.1: Data acquisition and 3D mesh generation.

[0066] Specifically, firstly, industrial cameras or AI vision cameras are deployed above automated assembly lines, material handling channels, workstation intersections, and buffer areas, in conjunction with structured light modules, binocular ranging modules, or laser ranging modules, to simultaneously acquire data from multiple perspectives within the workspace. The role of industrial vision is not merely to identify obstacles, but more importantly, to provide stable object categories, contour boundaries, movement trajectories, and workstation behavior characteristics for subsequent semantic partitioning. Subsequently, distortion correction, temporal alignment, and spatial calibration are performed on the images from each camera, mapping pixel-level target boundaries and depth values ​​to a unified spatial coordinate system. Then, the workspace is divided into a 3D mesh according to preset voxel edge lengths. The voxel edge length is 0.05-0.20 meters; for channel-type scenes, the voxel edge length is 0.08-0.12 meters to balance boundary accuracy and computational burden. Finally, the number of times each spatial unit is occupied, idle, and visible is recorded, forming a basic occupancy attribute table.

[0067] Sub-step 1.2, AI visual recognition and label writing.

[0068] Specifically, AI visual recognition is performed on industrial visual image sequences to identify object categories, workstation types, and behavioral states. Object categories include handling robots, robotic arm work areas, manual passage areas, buffer pallets, forklift aisles, and fixed equipment boundaries; area functions include main passage areas, meeting areas, workstation adjacency areas, waiting buffer zones, and temporary closed areas; behavioral states include continuous movement, brief stops, cross-flow, and periodic release. Furthermore, the visual model employs a cascaded structure of object detection and semantic segmentation. The confidence threshold for the object detection model is set at 0.55-0.80, and the region preservation threshold for the semantic segmentation model is set at 0.60-0.85. Subsequently, the recognition results are mapped onto three-dimensional spatial units, and the frequency of each spatial unit being covered by different semantic categories is counted. The category with the highest semantic classification score is selected as the primary semantic label for that spatial unit.

[0069] In this embodiment, the expression for the semantic classification score is:

[0070] ;

[0071] In the formula, Indicates the first Semantic classification score of each spatial unit; This indicates the average confidence level of the target detection corresponding to this spatial unit; This indicates the semantic segmentation coverage corresponding to the spatial unit; This indicates the number of times the spatial unit switched categories within the statistics window; This represents the detection confidence weighting index, with a value range of 1.0-2.0; This represents the segmentation coverage weight index, with a value range of 1.0-2.5; This represents the categorical volatility suppression index, with a value range of 0.5-1.5.

[0072] The expression for behavioral activity is:

[0073] ;

[0074] In the formula, Indicates the first The behavioral activity level of each spatial unit; This indicates the number of times a target traverses this spatial cell within the statistics window, expressed in times. This indicates the number of sampled frames within the statistical window length, in frames. and This represents the smoothing constant, with values ​​ranging from 0.1 to 1.0.

[0075] Sub-step 1.3: Semantic partition generation.

[0076] Specifically, semantic candidate units are aggregated into semantic partition sets based on spatial adjacency, consistency of primary semantic labels, and similarity of behavioral activity. During aggregation, adjacent spatial units are required to have consistent primary semantic labels, and the difference in behavioral activity must not exceed a set threshold, ranging from 0.10 to 0.30. For meeting areas and crossover / merging areas, additional manual correction rules or process-prior rules are needed to avoid mistakenly merging two adjacent but functionally different spatial regions into the same partition, ultimately resulting in a semantic partition set with clear regional boundaries and functional meanings.

[0077] Sub-step 1.4: Calculate the static and dynamic accessibility of the semantic partition.

[0078] Specifically, static and dynamic traffic capacity are calculated for each semantic partition. Static traffic capacity primarily reflects the basic throughput of the area when no moving objects are passing through, and is determined by the area's net height, net width, the proportion of fixed equipment intrusion, and the proportion of walkable ground area. Dynamic traffic capacity focuses on whether the semantic partition will periodically release traversable gaps under high-density collaborative operation conditions. For the main passage area, waiting buffer zone, and cycle release zone, industrial vision continuously monitors the flow cycle of moving objects, counts the number of actual gaps formed per unit time, and calculates the dynamic traffic capacity based on the gap duration. Finally, the static and dynamic traffic capacities are merged to form a traffic capacity model for each semantic partition.

[0079] In this embodiment, the expression for dynamic access capability is:

[0080] ;

[0081] In the formula, Indicates the first The dynamic throughput capacity of each semantic partition, in seconds; This indicates the number of times a passable gap is formed within the statistical window for this semantic partition; the unit is times. This represents the statistical time window, in seconds, with a range of 1-5 seconds.

[0082] The expression for static access capability is:

[0083] ;

[0084] In the formula, Indicates the first Static accessibility of each semantic partition; This represents the volume of free space that can be traversed within the semantic partition; This indicates the blocking volume within the semantic partition occupied by fixed equipment, guardrails, or workstation boundaries; This represents the volume smoothing constant, with a value ranging from 0.01 to 0.10 cubic meters.

[0085] Step S120: Based on the semantic partitioning accessibility model, perform spatial overlay analysis on the distribution of obstacles in the work space to identify closed areas and generate a semantic closed area distribution map.

[0086] In some embodiments, the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention includes the following steps in step S120:

[0087] Step S121: Continuously identify and track moving objects in the work space corresponding to each semantic partition to obtain the number, location and behavior status of the moving objects, and bind the number, location and behavior status of the moving objects with the dynamic accessibility to obtain the dynamic obstacle semantic coupling sequence.

[0088] Step S122: Perform time-series statistics on the semantic coupling sequence of dynamic obstacles to extract the instantaneous occupancy features of each semantic partition within a set time window, and calculate the closure index of the corresponding semantic partition based on the instantaneous occupancy features.

[0089] Step S123: Classify each semantic partition according to the degree of closure index to generate a semantic closure region distribution map.

[0090] In a specific embodiment, the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention includes step 2, which involves performing semantically driven dynamic closed region identification and hierarchical processing. Based on the semantic partition accessibility model output in step 1, spatial overlay analysis is performed on the distribution of dynamic obstacles to identify whether local areas form semantically temporary closed regions or high-frequency access areas. Subsequently, by calculating the ratio of instantaneous occupancy features to accessibility within the semantic partition, a closure degree index is obtained, and a semantic closed region distribution map is output, including the following sub-steps:

[0091] Sub-step 2.1: Construct a dynamic obstacle semantic coupling sequence.

[0092] Specifically, within the workspace corresponding to each semantic partition, industrial vision or AI vision systems are invoked to continuously identify, track, and semantically attach moving objects. In this example, industrial vision does not simply extract the location of obstacles, but rather identifies simultaneously "who the obstacle is, what its working state is, which semantic partition it is located in, and whether it forms an intersection or convergence." Therefore, a three-level vision processing structure combining object detection, multi-object tracking, and behavior recognition is adopted.

[0093] In this embodiment, target detection is used to identify mobile object categories such as handling robots, forklifts, manual carts, and pallet trucks; multi-target tracking is used to provide trajectory maintenance results for the same mobile object in consecutive frames; and behavior recognition is used to distinguish four types of behavior: straight-through, stopping and waiting, cutting into confluence, and lateral crossing. The target detection confidence threshold is 0.60-0.85, the tracking association threshold is 0.45-0.75, and the behavior duration frame threshold is 5-20 frames. Subsequently, the number of mobile objects identified at each moment, their occupied positions, motion states, and the dynamic passage capability of the corresponding semantic partition are bound together to form a dynamic obstacle semantic coupling sequence. Each record in the sequence includes at least the partition number, time number, number of obstacles, partition volume, dynamic passage capability, number of crossings, and stopping duration.

[0094] Sub-step 2.2: Extract the instantaneous occupancy features of the semantic partition.

[0095] Specifically, temporal statistics are performed on the semantic coupling sequence of dynamic obstacles to extract the instantaneous occupancy features, dominance of parking, and degree of cross-perturbation of each semantic partition within a given statistical window. In this example, the role of AI vision still needs to be emphasized, because whether a local area tends to be closed depends not only on the quantity, but also on whether the moving body is passing through normally in one direction, or crossing, repeatedly cutting in, or parking and waiting. If only the quantity is considered, fast-paced passage areas in the main channel may be misjudged as closed areas; only by introducing visual behavior recognition can the high-density but releaseable high-frequency passage areas be distinguished from high-density and difficult-to-release temporary closed areas. For the statistical window, a short window of 1-3 seconds is used; for high-paced assembly lines, it is shortened to 0.5-1.5 seconds.

[0096] In this embodiment, the expression for the instantaneous occupancy feature is:

[0097] ;

[0098] In the formula, Indicates the first Each semantic partition at time The instantaneous occupancy characteristics, expressed in units per cubic meter; Indicates the first Each semantic partition at time The number of dynamic obstacles confirmed by AI vision, in units of individuals; Indicates the first The spatial volume of each semantic partition is expressed in cubic meters. Indicates the first Each semantic partition at time The effective visibility coefficient, ranging from 0.55 to 1.00, is determined by a combination of the coverage of the industrial visual viewpoint and the occlusion ratio. This represents the volume smoothing constant, in cubic meters, with a value range of 0.01-0.10.

[0099] The expression for the degree of dominance of the station is:

[0100] ;

[0101] In the formula, Indicates the first The degree of dominance of each semantic partition; Indicated in the statistics window Inner The duration of a dynamic obstacle's stationary position, in seconds; This indicates the length of the statistical window, in seconds, and its value ranges from 1 to 5. Indicates the first The number of moving objects that are effectively traversed within a semantic partition within the statistics window, expressed in units of [number].

[0102] The expression for the degree of cross disturbance is:

[0103] ;

[0104] In the formula, Indicates the first The degree of cross-perturbation of each semantic partition; This indicates the number of crossover events identified by AI vision within the statistics window, expressed in times. This indicates the number of events that pass normally along the main direction within the statistics window, expressed in times. This represents the event smoothing constant, with a value range of 0.5-2.0.

[0105] Sub-step 2.3: Calculate the semantic closure index.

[0106] Specifically, a closure index is calculated for each semantic partition, and the instantaneous occupancy status is normalized based on the dynamic accessibility of the semantic partition. It should be noted that while the raw ratio provided by the user reflects the basic idea that "higher occupancy density and lower accessibility indicate greater closure," a normalization process is necessary to ensure consistency of units. First, the instantaneous occupancy characteristics and dynamic accessibility are normalized to a baseline, and then the ratio is calculated. The normalization baseline value is obtained from statistics collected during the debugging phase under normal production line cycle time. If the production line cycle time is stable, the density baseline value is 2-8 cubic meters per second, and the accessibility baseline value is 0.2-2.0 cubic meters per second.

[0107] In this embodiment, the expression for the degree of closure index is:

[0108] ;

[0109] In the formula, Indicates the first The degree of closure of a semantic partition; This represents the density normalized value; Indicates the dominance coefficient of the stationing; Indicates the cross disturbance coefficient; This represents the normalized value of the traffic capacity; This represents the dwell time amplification index, with a value range of 0.8-1.8; This represents the cross-amplification index, with a value range of 1.0-2.5; It represents a very small positive number, with a value range of 0.01-0.10.

[0110] Sub-step 2.4 generates a semantically closed region distribution map.

[0111] Specifically, based on the closure level index, each semantic partition is graded and a semantic closure region distribution map is generated. A three- or four-level grading method is adopted. Taking the three-level method as an example, when the closure level index is less than the first threshold, it is marked as a high-frequency passage area; when the closure level index is between the first and second thresholds, it is marked as a moderately congested area; and when the closure level index is greater than the second threshold, it is marked as a temporary closed area. The first threshold ranges from 0.8 to 1.5, and the second threshold ranges from 1.5 to 3.5. To prevent the grading results from fluctuating between adjacent time points, a time-preservation rule is further introduced, meaning that the same semantic partition must meet the threshold condition for two to three consecutive statistical windows before the state switch is completed. Finally, the grading results are superimposed on the three-dimensional semantic partitioning model to form a semantic closure region distribution map with descriptions of spatial location, closure level, dominant behavior type, and passage capability.

[0112] Step S130: The robot's initial path is segmented according to the semantic closed region distribution map, and a semantic access window is constructed within the closed region to generate a candidate path sequence constrained by the semantic access window.

[0113] In some embodiments, the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention includes the following steps in step S130:

[0114] Step S131: Map the robot's initial path segment by segment to the semantic closed region distribution map, and divide each path segment into a passable segment and a semantically restricted segment.

[0115] Step S132: Determine the semantic access window corresponding to the semantically restricted segment based on the dynamic access capacity and closure degree index, and construct multiple candidate access paths based on the semantic access window.

[0116] Step S133: The candidate travel paths and travel segments are spliced ​​together according to the travel order of each segment in the robot's initial path, and the spliced ​​travel paths are sorted and trimmed to obtain the candidate path sequence.

[0117] In a specific embodiment, the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention includes step 3, which involves performing path candidate reconstruction processing based on semantic access windows. Based on the semantic closed region distribution map output in step 2, the original path is segmented. Within the closed region, obstacle avoidance is no longer performed directly; instead, temporal and spatial access windows are generated based on semantic accessibility, and a candidate path set is constructed. The result is an output set of candidate path sequences under semantic constraints, including the following sub-steps:

[0118] Sub-step 3.1: Initial path segmentation for the robot.

[0119] Specifically, firstly, the original path sequence is mapped onto a semantically closed region distribution map, and each segment is judged to determine whether it traverses a temporary closed region, a moderately congested region, or a high-frequency passage region. In this example, instead of simply bypassing obstacles as in existing technologies, the original path is first divided into stable passage segments and semantically restricted segments. Stable passage segments refer to paths traversing areas that are high-frequency passage regions within the current statistical window, and whose closure level is below the first judgment threshold. Semantically restricted segments refer to paths entering moderately congested or temporary closed regions, requiring subsequent construction of temporal and spatial passage windows based on semantic passage capabilities. Path segmentation is performed on continuous grid segments. When five or more consecutive path nodes fall into the same semantic level region, they are merged into a candidate path segment. If a path segment crosses different semantic level regions, the boundary of the level change is used as the segmentation point. After the above processing, the robot's original path is decomposed into several interconnected sub-path units, each with a clear semantic label, closure level, path length, and expected direction of travel.

[0120] Sub-step 3.2: Calculate the time window in which passage gaps can be formed in the work area in the future.

[0121] Specifically, for each semantically restricted segment, instead of performing immediate obstacle avoidance based on the current position, the dynamic passage capability from step 1 and the degree of closure from step 2 are used to calculate the time window in which passage gaps can be formed in the current area in the future. In this example, the time and space passage window includes three core components: first, the window start waiting time, indicating when the robot should enter the restricted segment at the earliest; second, the window duration, indicating the length of time during which the restricted segment allows continuous passage; and third, the window adaptation width, indicating the effective width of the restricted segment that can be passed through within the time window.

[0122] It should be noted that if the unit of dynamic throughput capacity is per second, the window duration cannot be simply written as a product of it, otherwise the dimensions would be invalid. Therefore, the reciprocal of the throughput capacity is used to construct the time quantity. For automated transport channels with stable cycle times, the greater the throughput capacity of the high-frequency release zone, the shorter the average waiting time, not the longer the window time.

[0123] In this embodiment, the expression for the duration of a single passage window for a semantically restricted segment is:

[0124] ;

[0125] In the formula, Indicates the first The duration of a single passage window for a semantically restricted segment, in seconds; This represents the time amplification constant, expressed in seconds, and its value ranges from 0.5 to 2.0. Indicates the first The dynamic throughput capacity of the semantic partition to which a semantically restricted segment belongs, in seconds; This represents the smoothing constant for traffic capacity, measured in seconds, and ranging from 0.01 to 0.10. Indicates the first An indicator of the degree of closure of the semantic partition to which a semantically restricted segment belongs; This represents the closed decay index, with a value range of 0.5-1.5.

[0126] The expression for the expected initial latency of a semantically restricted segment is:

[0127] ;

[0128] In the formula, Indicates the first The estimated start-up wait time for each semantically restricted segment, in seconds; Indicators representing the degree of closure; This indicates the amplification index, with a value range of 0.8-2.0; Indicates dynamic throughput capacity, measured in seconds; This represents the smoothing constant for traffic capacity, measured in seconds, and ranging from 0.01 to 0.10.

[0129] The expression for the effective passage width of a semantically restricted segment within the passage window is:

[0130] ;

[0131] In the formula, Indicates the first The effective passage width of a semantically restricted segment within the passage window, in meters; Indicates the first The static net width of the region where each semantically restricted segment is located, in meters, is obtained from the semantic partitioning geometric attributes in step 1. Indicators representing the degree of closure; This represents the width compression index, with a value range of 0.3-1.2.

[0132] Sub-step 3.3: Construct candidate passable sub-paths.

[0133] Specifically, for each semantically restricted segment, three types of candidate passage sub-paths are constructed around its passage window: the first type is the delayed direct passage type, which maintains the original path direction unchanged, enters the window and passes quickly after the expected initial waiting time arrives; the second type is the boundary fitting type, which shifts the path locally to the low-disturbance boundary side of the semantically restricted segment when the effective passage width is compressed, so as to avoid the intersection and merging center; the third type is the short-distance bypass type, which only bypasses the strongest closing kernel inside the semantically closed area, without bypassing the entire partition, thereby avoiding large-scale backtracking.

[0134] When generating candidate passage sub-paths, based on the closed area distribution map output in step 2, the highly closed core area and the relatively buffer edge area within the restricted section are identified. The highly closed core area generally corresponds to the intersection of multiple moving bodies or the location of concentrated parking; the buffer edge area corresponds to the area with relatively high traffic capacity and a large effective passage width. For each type of candidate sub-path, the corresponding entry time, passage time, sub-path length, minimum clearance, and expected waiting time need to be generated. Therefore, a restricted heuristic search or local graph search method is adopted, with a search radius of 1.2-2.0 times the length of the restricted section, and the number of candidate sub-paths is 3-8. This is because too few sub-paths will lead to insufficient options, while too many sub-paths will increase the burden of subsequent combination.

[0135] Sub-step 3.4 generates a sequence of candidate paths under semantic constraints.

[0136] Specifically, the candidate traversable sub-paths of each semantically restricted segment are concatenated with the stable traversable segments retained in step 3.1 in their original sequential order to form multiple complete candidate path sequences. During concatenation, geometric connectivity alone is insufficient; temporal connectivity and semantic consistency must also be checked. Temporal connectivity means that the entry time of the subsequent sub-segment must not be earlier than the departure time of the preceding sub-segment; semantic consistency means that there should be no aberrational switching at the concatenation boundary between adjacent sub-segments, where the sub-segment jumps directly from a low-traffic-capacity region to a high-closure-core region. Furthermore, to reduce candidate explosion, filtering must be performed immediately after each concatenation, retaining the top 5-15 complete path sequences with the lowest total cost.

[0137] For each complete candidate path sequence, at least the total path length, total time, minimum net distance, average closure contact rate, and expected straight-through success rate should be calculated. The expected straight-through success rate is obtained by comprehensively considering the window utilization sufficiency of all semantically restricted segments traversed by the path sequence. Higher window utilization sufficiency indicates a better match between the path and the release timing of each restricted segment, leading to a higher probability of subsequent successful execution.

[0138] Sub-step 3.5 outputs the candidate path sequence under semantic constraints.

[0139] Specifically, the concatenated candidate path sequences undergo final sorting and pruning, eliminating path sequences whose length increases excessively compared to the original path, whose total time exceeds the allowable threshold, or whose expected direct success rate is too low. The length increase threshold is 15%-40%, the upper limit of total time is 1.2-2.0 times the expected time of the original path, and the lower limit of success rate is 0.50-0.80. After screening, several executable candidate path sequences are retained.

[0140] Step S140: Sort the candidate path sequences by semantic priority factor to obtain the optimal path sequence guided by semantics.

[0141] In some embodiments, the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention includes the following steps in step S140:

[0142] Step S141: Semantic priority expansion is performed on the semantic partitions where each path segment in the candidate path sequence is located to construct a path segment evaluation table corresponding to the candidate path sequence, and a semantic priority factor is constructed based on the path segment evaluation table.

[0143] Generally, the path length can be converted into travel time based on the closure level index, dynamic capacity, window duration, and estimated waiting time of each path segment in the path segment evaluation table. The closure level can be mapped into additional waiting time, and then combined with the reciprocal of the dynamic capacity to form a passage opportunity waiting item. After being uniformly converted into time cost, a weighted fusion is performed to obtain the semantic priority factor. For example, if a path segment has a low closure level and a high capacity, its overall time cost is lower, and its priority is higher.

[0144] Step S142: Sort the candidate path sequences according to the semantic priority factor to obtain the semantic priority ranking result of the candidate path sequences, and select the candidate path sequence with the highest ranking from the semantic priority ranking result of the candidate path sequences as the initial optimal path sequence.

[0145] In some embodiments, the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention further includes the following steps in step S140:

[0146] Step S143: When executing the initial optimal path sequence, the updated semantic closed region distribution map is received in real time, and the path segment to be executed is reviewed to determine whether the path segment to be executed meets the constraints of the semantic passage window.

[0147] Step S144: When the path segment to be executed does not meet the constraints of the semantic passage window, select the path segment with the highest semantic priority from the candidate path sequence to replace the path segment to be executed, and obtain the optimal path sequence guided by semantics.

[0148] In a specific embodiment, the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention includes step 4, which involves performing semantic priority-guided path selection and dynamic adjustment processing. Based on the candidate path sequence set output in step 3, a semantic priority factor is introduced to sort the paths, prioritizing path segments with high traffic capacity and low closure. During execution, the paths are locally replaced and spliced ​​according to real-time semantic partition changes, outputting a semantically guided dynamic optimal path sequence, including the following sub-steps:

[0149] Sub-step 4.1, candidate path sequence decomposition.

[0150] Specifically, firstly, each candidate path sequence output in step 3 is uniformly decomposed, extracting the total path length, the number of semantic partitions traversed by the path, the closure degree corresponding to each semantic partition, dynamic accessibility, window success rate, and the number of path segment switching times. This decomposition process does not regenerate the path, but rather transforms the candidate path sequence into a comparable semantic evaluation object. Since step 3 has already reconstructed the original path into several candidate path sequences with time and space access window constraints, it is necessary to perform semantic priority expansion on the semantic partition where each path segment belongs, following the execution order of the path sequence. Here, semantic priority is a quantitative expression of whether a path segment is more worthy of priority.

[0151] In this embodiment, a path segment-level evaluation table needs to be established for each candidate path sequence. This evaluation table includes at least the path segment number, path segment length, semantic partition number, closure index, dynamic traffic capacity, window duration, estimated waiting time, and window matching degree. For multiple path segments within the same path sequence, if the path segment length is too short, for example, less than 0.3 meters, adjacent semantic segments of the same type are merged first to reduce subsequent calculation oscillations. For scenarios with a large number of candidate paths, the top 10-20 paths with the highest comprehensive cost are retained first, and then fine semantic sorting is performed to balance real-time performance and accuracy.

[0152] Sub-step 4.2: Sort the candidate path sequence.

[0153] Specifically, based on the semantic evaluation parameter set of candidate path sequences, a semantic priority factor is constructed, and all candidate path sequences are sorted. In this example, the original path evaluation function provided by the user includes three terms: path length, degree of closure, and the reciprocal of dynamic throughput. While their engineering meaning is correct, directly adding these three terms requires first unifying the units through a conversion factor; otherwise, inconsistencies in units will occur. Therefore, the evaluation function needs to be constructed by converting all terms into time costs.

[0154] Specifically, the path length is converted into travel time using reference speed, the degree of closure is converted into additional waiting time using risk waiting mapping, and the reciprocal of dynamic throughput capacity is directly retained as a throughput opportunity waiting item. The resulting comprehensive evaluation is uniformly measured in seconds for easy sorting. Furthermore, regarding the weighting of the sorting criteria, if the scenario prioritizes high-speed passage, the throughput capacity item weight is appropriately increased; if the scenario prioritizes avoiding lock-up, the closure degree item weight is appropriately increased; if the scenario prioritizes the shortest transport distance, the path length item weight is increased. The weights for these three items are set between 0.2 and 0.6, and their sum is 1. For highly dynamic scenarios such as assembly line junctions, the closure degree item weight is set at 0.35-0.50, the throughput capacity item weight at 0.30-0.45, and the path length item weight at 0.20-0.35.

[0155] Sub-step 4.3: Select the initial optimal path sequence.

[0156] Specifically, after sorting the candidate path sequences by comprehensive evaluation cost from smallest to largest, the top-ranked candidate path sequence is selected as the initial optimal path sequence. Simultaneously, to prevent local mismatches from occurring before actual execution, a backup path pool needs to be established for the top 3-5 candidate path sequences. The path sequences in this backup pool should have differences in key semantically restricted segments, rather than just minor length differences. In this example, if multiple candidate paths rely on the same temporary closed area release, they should not all be retained as backup paths; paths relying on different semantic windows, different boundary alignment directions, or different local bypass methods should be prioritized to enhance the replaceability of subsequent dynamic adjustments.

[0157] For the initial optimal path sequence, an execution schedule must also be generated synchronously. This execution schedule should include at least the expected entry time, expected exit time, and corresponding semantic window matching interval for each path segment. If the expected entry time of a path segment is later than its window closing time, it should fall back to the next candidate path in the ranking results for re-verification.

[0158] Sub-step 4.4: Construct local replacement trigger flags and local replacement candidate segments.

[0159] Specifically, during the robot's execution along the initial optimal path sequence, it continuously receives real-time semantic partition change information obtained from the link update in step 2, and performs rolling verification on the path segments within a certain range ahead of the current path. The rolling verification range is 1-5 meters in the robot's current direction of movement or the path segment expected to be entered within the next 2-8 seconds. If any of the following conditions are detected in a critical path segment in the current path, a local replacement is triggered: first, the closure degree exceeds the preset upper limit for two consecutive statistical windows; second, the dynamic passage capability drops below the preset lower limit; third, the duration of the previously available passage window is shortened to be insufficient to support the robot's passage.

[0160] In this embodiment, once a local replacement is triggered, the entire path is not discarded. Instead, only the local path interval containing the affected path segment is extracted, and alternative segments that are connected to the start and end points of this interval are extracted from the backup path pool. This local replacement process uses the candidate path segments generated in step 3 for local patching, rather than global replanning, which can significantly reduce the frequent rollback problem in high-density collaborative scenarios.

[0161] Sub-step 4.5 outputs the semantically guided dynamic optimal path sequence.

[0162] Specifically, when a local replacement is triggered, the replacement segment with the lowest connection cost and highest semantic priority to the current path interval is selected from the local replacement candidate segment set. This replacement segment is then reassembled with the unaffected preceding and following path segments to form an updated dynamic optimal path sequence. If no local replacement is triggered, the initial optimal path sequence is maintained and execution continues. After the assembly is completed, the temporal connectivity, window matching degree, and minimum clearance of the overall path must be verified again to ensure that the updated dynamic optimal path sequence remains continuously executable. Finally, the output dynamic optimal path sequence should include at least the path point sequence, segment semantic labels, the expected entry time of each key path segment, the matching results of each key window, and the current comprehensive evaluation cost.

[0163] Step S150: When a path segment in the optimal path sequence enters a closed region, the path segment is delayed or reconstructed through a semantic access window to output a path planning result that satisfies the semantic access window constraints.

[0164] In some embodiments, the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention includes the following steps in step S150:

[0165] Step S151: Perform continuity detection on the execution order of each path segment in the optimal path sequence to obtain continuity detection results, and perform risk screening on the continuity detection results to identify risky path segments that enter the closed area.

[0166] Step S152: Determine whether the risk path segment has a semantic access window. If the risk path segment has a semantic access window, perform delayed execution processing on the risk path segment; otherwise, replace the risk path segment.

[0167] The path planning results include risky path segments that have been delayed or replaced.

[0168] In a specific embodiment, the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention includes step 5, which involves performing path continuity assurance and anti-lock-up output processing. Based on the dynamic optimal path sequence output in step 4, path continuity detection is performed. When a path segment is detected to enter a highly enclosed semantic region, a semantic passage window mechanism is triggered for delayed passage or local replanning, ultimately generating an executable path result without path lock-up and outputting the final path planning result. This includes the following sub-steps:

[0169] Sub-step 5.1: Detect the continuity of the execution order of each path segment in the dynamic optimal path sequence.

[0170] Specifically, firstly, continuity checks are performed on each path segment in the dynamic optimal path sequence according to the execution order. The checks include at least spatial continuity, temporal continuity, and semantic continuity. Spatial continuity is used to determine whether there are abrupt transitions, abnormal jumps, or sudden drops in clearance between adjacent path segments; temporal continuity is used to determine whether the departure time of the previous passage window conflicts with the entry time of the next passage window; and semantic continuity is used to determine whether the path suddenly cuts from a low-closure region to a high-closure region without reserving a waiting buffer. For each path segment, its corresponding path length, expected entry time, expected departure time, semantic closure degree, and dynamic passage capability are read, and a continuity check result set is formed. For high-speed transport robots, the continuity check period ranges from 0.1 to 0.5 seconds; for low-speed collaborative robots, the continuity check period is 0.3 to 1.0 seconds.

[0171] Sub-step 5.2, path segment risk screening.

[0172] Specifically, the path continuity detection result set is further subjected to risk screening, with a focus on identifying path segments that are about to enter highly enclosed semantic regions. When the enclosure level of a path segment continuously increases, its dynamic passage capacity continuously decreases, and the current time buffer is insufficient to ensure that the robot can complete the passage within the window, a high-enclosure risk trigger marker is generated. In the identification process, the judgment is not based solely on a single enclosure level value, but rather on a joint judgment combining the remaining window time, the theoretical passage time of the path segment, and the extent of local semantic deterioration. For automated assembly lines with large cycle time fluctuations, the risk trigger threshold should be moderately conservative; for scenarios with wider passages and the ability to move and yield, the threshold can be moderately relaxed.

[0173] Sub-step 5.3 generates delayed decision results or local replanning decision results.

[0174] Specifically, when a path segment is triggered as a high-risk closed path segment, the system first determines whether there is still a usable semantic passage window for that path segment. If so, a delayed passage decision is executed, meaning that while the current path remains unchanged, the entry time is adjusted, and entry is delayed until the next valid window opens. If no window exists, or the window length is insufficient to cover the theoretical passage time, a local replanning decision is executed, selecting a replacement segment from the backup path segments formed in step 4 and locally splicing them at the boundaries before and after the current path. Furthermore, to avoid frequent strategy switching, a minimum waiting time threshold and a minimum replacement benefit threshold need to be set. The minimum waiting time threshold ranges from 0.3 to 3.0 seconds, and the minimum replacement benefit threshold ranges from 5% to 20% of the time reduction.

[0175] Sub-step 5.4 outputs the final path planning result.

[0176] Specifically, if a delayed passage decision is made, the estimated entry and exit times of the corresponding path segment and the timetable of subsequent path segments are updated. If a local replanning decision is made, the original high-risk path segment is replaced with an alternative path segment, and the spatial continuity, temporal continuity, and minimum clearance of adjacent path segments are re-verified. After the update is completed, the final path planning result is output. This final path planning result includes at least the path point sequence, semantic labels of each path segment, the estimated entry time of each key path segment, the final window matching result, and the current lock-up risk status. This final path planning result is then directly sent to the robot control terminal for execution.

[0177] The robot optimal path planning device based on 3D environmental semantics and dynamic obstacle avoidance provided by the present invention will be described below. The robot optimal path planning device based on 3D environmental semantics and dynamic obstacle avoidance described below can be referred to in correspondence with the robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance described above.

[0178] like Figure 2 As shown in one embodiment, a robot optimal path planning device based on 3D environmental semantics and dynamic obstacle avoidance includes a traffic model construction module, a closed area identification module, a candidate path generation module, an optimal path screening module, and a path planning output module.

[0179] The access model construction module is used to acquire workspace perception data and perform three-dimensional mesh division and semantic classification on the workspace perception data in order to construct a semantic partition access capability model.

[0180] The closed area identification module is used to perform spatial overlay analysis on the distribution of obstacles in the work space based on the semantic partitioning accessibility model, so as to identify closed areas and generate a semantic closed area distribution map.

[0181] The candidate path generation module is used to segment the robot's initial path according to the semantic closed region distribution map, and construct a semantic access window within the closed region to generate a candidate path sequence constrained by the semantic access window.

[0182] The optimal path selection module is used to sort candidate path sequences by semantic priority factors in order to obtain the semantically guided optimal path sequence.

[0183] The path planning output module is used to delay or reconstruct path segments in the optimal path sequence when they enter a closed region, so as to output path planning results that satisfy the semantic passage window constraints.

[0184] The applicant of this invention has provided a detailed description of the embodiments of the invention in conjunction with the accompanying drawings. However, those skilled in the art should understand that the above embodiments are merely preferred embodiments of the invention. The detailed description is only intended to help readers better understand the spirit of the invention and is not intended to limit the scope of protection of the invention. On the contrary, any improvements or modifications made based on the inventive spirit of the invention should fall within the scope of protection of the invention.

Claims

1. A robot optimal path planning method based on 3D environmental semantics and dynamic obstacle avoidance, characterized in that, The method includes: Acquire workspace perception data, and perform three-dimensional mesh division and semantic classification on the workspace perception data to construct a semantic partition accessibility model; Based on the semantic partitioning accessibility model, spatial overlay analysis is performed on the distribution of obstacles in the work space to identify closed areas and generate a semantic closed area distribution map. The robot's initial path is segmented according to the semantic closed region distribution map, and a semantic access window is constructed within the closed region to generate a candidate path sequence constrained by the semantic access window. The candidate path sequences are sorted by a semantic priority factor to obtain the optimal semantically guided path sequence. When a path segment in the optimal path sequence enters a closed region, the path segment is delayed or reconstructed through the semantic access window to output a path planning result that satisfies the semantic access window constraints.

2. The robot optimal path planning method based on 3D environment semantics and dynamic obstacle avoidance according to claim 1, characterized in that, The process of acquiring operational space perception data and performing three-dimensional mesh partitioning and semantic classification on the operational space perception data to construct a semantic partitioning accessibility model includes: Visual images and depth ranging data of the workspace are acquired through image acquisition equipment and a ranging module, and distortion correction and alignment calibration are performed on the visual images and depth ranging data. The work space is divided into three-dimensional meshes according to the preset voxel side lengths, and the object category, region function and behavior state corresponding to the visual image are identified by the visual model so as to map the object category, region function and behavior state to the corresponding spatial units respectively. The spatial unit is obtained by dividing the work space into three-dimensional meshes, the visual model is a cascaded structure of a target detection model and a semantic segmentation model, and the work space perception data includes the visual image and depth ranging data.

3. The robot optimal path planning method based on 3D environment semantics and dynamic obstacle avoidance according to claim 2, characterized in that, The process of acquiring workspace perception data and performing three-dimensional mesh partitioning and semantic classification on the workspace perception data to construct a semantic partitioning accessibility model further includes: The semantic scores of different semantic categories in each spatial unit are statistically analyzed. The semantic category with the highest semantic score is selected as the semantic label of the corresponding spatial unit. The spatial units are then aggregated based on the adjacency relationship between different spatial units and the similarity between semantic labels and behavioral states to obtain multiple semantic partitions. Calculate the static access capability and dynamic access capability corresponding to each semantic partition, and merge the static access capability and dynamic access capability into the corresponding semantic partition to obtain the semantic partition access capability model.

4. The robot optimal path planning method based on 3D environment semantics and dynamic obstacle avoidance according to claim 3, characterized in that, The step of performing spatial overlay analysis on the obstacle distribution within the work space based on the semantic partitioning accessibility model to identify enclosed areas and generate a semantic enclosed area distribution map includes: The moving objects in the work space corresponding to each semantic partition are continuously identified and tracked to obtain the number, location and behavior status of the moving objects, and the number, location and behavior status of the moving objects are bound to the dynamic accessibility to obtain a dynamic obstacle semantic coupling sequence. Temporal statistics are performed on the dynamic obstacle semantic coupling sequence to extract the instantaneous occupancy features of each semantic partition within a set time window, and the closure index of the corresponding semantic partition is calculated based on the instantaneous occupancy features. The semantic partitions are classified according to the closure index to generate the semantic closure region distribution map.

5. The robot optimal path planning method based on 3D environment semantics and dynamic obstacle avoidance according to claim 4, characterized in that, The step of segmenting the robot's initial path according to the semantic closed region distribution map and constructing a semantic access window within the closed region to generate a candidate path sequence constrained by the semantic access window includes: The robot's initial path is mapped segment by segment to the semantically closed region distribution map, and each path segment is divided into a passable segment and a semantically restricted segment; The semantic access window corresponding to the semantically restricted segment is determined based on the dynamic access capability and closure degree index, and multiple candidate access paths are constructed based on the semantic access window; The candidate paths and segments are spliced ​​together according to the passage order of each segment in the robot's initial path, and the spliced ​​paths are sorted and trimmed to obtain the candidate path sequence.

6. The robot optimal path planning method based on 3D environment semantics and dynamic obstacle avoidance according to claim 1, characterized in that, The step of sorting the candidate path sequences using a semantic priority factor to obtain the optimal semantically guided path sequence includes: Semantic priority expansion is performed on the semantic partitions where each path segment in the candidate path sequence is located to construct a path segment evaluation table corresponding to the candidate path sequence, and the semantic priority factor is constructed based on the path segment evaluation table; The candidate path sequences are sorted according to the semantic priority factor to obtain the semantic priority ranking result of the candidate path sequences, and the candidate path sequence with the highest ranking is selected from the semantic priority ranking result of the candidate path sequences as the initial optimal path sequence.

7. The robot optimal path planning method based on 3D environment semantics and dynamic obstacle avoidance according to claim 6, characterized in that, The step of sorting the candidate path sequences using a semantic priority factor to obtain the optimal semantically guided path sequence further includes: When executing the initial optimal path sequence, the updated semantic closed region distribution map is received in real time, and the path segment to be executed is reviewed to determine whether the path segment to be executed meets the constraints of the semantic passage window. When the path segment to be executed does not meet the constraints of the semantic passage window, the path segment with the highest semantic priority is selected from the candidate path sequence to replace the path segment to be executed, thereby obtaining the optimal path sequence guided by semantics.

8. The robot optimal path planning method based on 3D environment semantics and dynamic obstacle avoidance according to claim 1, characterized in that, When a path segment in the optimal path sequence enters a closed region, the path segment is delayed or reconstructed through the semantic access window to output a path planning result that satisfies the semantic access window constraints, including: The execution order of each path segment in the optimal path sequence is continuously checked to obtain the continuous check results. The continuous check results are then used for risk screening to identify risky path segments that enter the closed area. Determine whether the risk path segment has a semantic access window, and if the risk path segment has the semantic access window, perform delayed execution processing on the risk path segment; otherwise, replace the risk path segment. The path planning results include risky path segments that have undergone delayed execution processing or replacement.

9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-8.

Images