A semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots

Abstract

This invention provides a semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots, relating to the field of autonomous navigation technology for mobile robots. The method includes: assigning semantic labels to each grid in a grid image of the scene in which the target robot is located; calculating an initial confidence score for the semantic labels based on a multi-source observation set; calculating the semantic reliability of the semantic labels based on the initial confidence score; constructing a cost function for the target robot; and performing global path planning on the target robot using a global path search algorithm based on the cost function to obtain a global path skeleton; generating the execution actions of the target robot; and generating the global path skeleton and execution actions based on the updated semantic reliability. This application improves the scene adaptability and safety of robot path planning, enhances the stability of local dynamic obstacle avoidance, and achieves adaptive closed-loop adjustment of navigation decisions.

Description

Technical Field

[0001] This invention relates to the field of autonomous navigation technology for mobile robots, and in particular to a semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots. Background Technology

[0002] Quadruped robot technologies and products are typically used in scenarios such as parks, campuses, warehouses, and logistics distribution. These systems enable robots to autonomously generate travel paths based on their starting point, destination, and surrounding environment, and to perform localization, obstacle avoidance, and motion control during movement. Such systems generally consist of a sensor perception unit, a map building unit, a path planning unit, and a motion execution unit. Sensors collect environmental and robot state information, the map building unit generates a map representation suitable for planning, the path planning unit generates a path from the starting point to the destination, and the motion execution unit controls the robot's movement according to the planning results. Currently, the autonomous navigation technology for quadruped robots mainly suffers from the following technical shortcomings: 1. The environmental representation focuses on geometric accessibility, but lacks sufficient characterization of scene attributes. Existing path planning methods for quadruped robots typically rely on terrain information such as elevation maps, slope, curvature, and roughness to construct a accessibility cost map. These methods then incorporate robot kinematics and dynamic constraints for global and local path planning. This approach is suitable for solving the problem of quadruped robots navigating complex terrains such as stairs, steps, and ditches. Its core criteria are terrain structure and robot mobility.

[0003] However, in semi-open environments such as campuses, industrial parks, and residential areas, robot accessibility is not entirely determined by geometric terrain. While some areas may appear passable based on topography, their actual accessibility is influenced by factors such as area purpose, access rules, crowd activity, and temporary obstacles. If environmental representation primarily focuses on geometric dimensions like elevation, slope, and roughness, it fails to fully reflect the accessibility conditions in real-world scenarios.

[0004] 2. The coupling between semantic understanding results and underlying navigation control is limited. Existing navigation methods for automated guided vehicles (AGVs) based on visual language models can generate scene description information from cargo instructions and scene images, and combine this information with cargo area, road information, and obstacle information to generate navigation paths. These methods enhance the robot's ability to understand task objectives and work areas, and are suitable for tasks such as target cargo identification, cargo area localization, and path generation in logistics handling scenarios.

[0005] However, from the perspective of autonomous navigation in complex environments, the semantic information in these methods is mainly used for scene description, target area determination, and navigation path generation. Its focus remains on helping the robot understand "where the target is" and "how to get near the target." Existing solutions have relatively limited ability to further transform semantic information into traffic constraints, risk costs, local obstacle avoidance criteria, or motion control constraints in the underlying navigation process. Therefore, in scenarios such as campuses or parks where humans and robots coexist, there are many rule constraints, and traffic conditions change rapidly, relying solely on this type of semantic understanding is insufficient to fully support the robot's fine-grained decision-making in complex traffic environments.

[0006] 3. Limited adaptability to dynamically changing scenarios. Existing path planning methods typically rely on pre-built maps, terrain costs, or scene perception results at a given moment for path generation. For relatively stable environments such as warehouses, fixed passageways, or known terrain, these methods can perform navigation tasks well. Quadruped robot accessibility planning schemes mainly revolve around elevation maps and terrain features, while visual language model navigation schemes primarily generate paths based on cargo areas, road information, and obstacle information.

[0007] However, in campus or industrial park logistics scenarios, factors such as pedestrian traffic, vehicle parking, shared bicycle accumulation, temporary construction barriers, and building entrance congestion can change over time. If the system mainly relies on static or short-term perception results for planning, it may be difficult to reflect changes in environmental conditions in a timely manner, leading to situations such as detour difficulties, reduced traffic efficiency, or the need for frequent replanning during the execution of the generated path.

[0008] 4. Insufficient feedback between global planning and local execution Existing quadruped robot path planning schemes typically first construct a accessibility cost map, then generate a global path, and finally select local target points on the global path for real-time local planning. Visual language model navigation schemes, on the other hand, usually first generate a scene description based on instructions and images, and then generate a path leading to the target area or the vicinity of the target cargo. These processes have a relatively clear sequence and can complete basic path generation and execution tasks.

[0009] However, in actual operation, situations that cannot be fully foreseen in the global planning phase may occur during the local execution phase. These include sudden appearances of local obstacles, temporary blockages in local pathways, and the robot repeatedly avoiding obstacles or briefly pausing in a certain area. Existing solutions emphasize the "plan first, execute later" process, and do not adequately address how the results of local execution can correct existing maps, paths, or environmental judgments. Therefore, the system's adaptive adjustment capabilities still have room for improvement when local execution deviates from the original plan. Summary of the Invention

[0010] This invention provides a semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots, aiming to solve the problems in existing technologies, such as environmental representation focusing on geometric information while ignoring scene attributes, insufficient coupling between semantic understanding and low-level control, poor adaptability to dynamic scenes, and lack of effective feedback between global planning and local execution.

[0011] In a first aspect, the present invention provides a semantically constrained path planning and dynamic obstacle avoidance method for a quadruped robot, comprising: Acquire a multi-source observation set, which includes the target robot's state information, laser point cloud data collected by the target robot, and visual observation data; Assign a semantic label to each grid in the grid image of the scene in which the target robot is located; calculate an initial confidence score for the semantic label based on the multi-source observation set; wherein, the semantic label system includes multiple semantic labels, and the semantic labels are used to characterize the semantic category of the semantic information of the scene in which the target robot is located; The semantic reliability of the semantic tag is calculated based on the initial confidence score, and the semantic reliability is used to characterize the degree to which the semantic tag is trusted. A cost function for the target robot is constructed based on the semantic reliability. Based on the cost function, a global path search algorithm is used to perform global path planning for the target robot to obtain a global path skeleton. The global path skeleton is used to guide the target robot from the task start point to the task end point. The action to be performed by the target robot during its movement along the global path skeleton is generated based on the policy network; Determine whether the target robot's action was successfully executed. If the determination result is negative, update the semantic reliability and generate a global path skeleton and action based on the updated semantic reliability.

[0012] In a second aspect, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the semantic constraint path planning and dynamic obstacle avoidance method for quadruped robots described above.

[0013] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots as described above.

[0014] Fourthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the semantic constraint path planning and dynamic obstacle avoidance method for quadruped robots described above.

[0015] According to the specific embodiments provided in this application, the following technical effects are disclosed: 1. This invention constructs a multi-dimensional semantic tagging system encompassing terrain, rules, dynamics, and tasks, and quantifies this system as semantic reliability, incorporating it into the cost function of global path planning. This enables the robot not only to recognize geometric obstacles but also to understand and abide by the social attributes and rules of passage within the scene. This effectively prevents the robot from entering areas that are "geometrically passable but actually impassable," making the planned paths safer, more reasonable, and in line with social expectations.

[0016] 2. This invention employs a temporal reinforcement learning model capable of processing historical information for local obstacle avoidance. The robot can infer the movement trends of dynamic obstacles such as pedestrians and non-motorized vehicles based on continuous observation sequences. Compared to reactive obstacle avoidance that relies solely on single-frame observations, this method can make more predictive decisions, effectively reducing hesitation, sudden stops, and frequent swaying in human-robot mixed scenarios, and improving the stability and smoothness of local passage.

[0017] 3. This invention establishes a feedback and update mechanism between global planning and local execution, enabling the system to transform unexpected situations encountered during local execution (such as path blocking or obstacle avoidance failure) into corrections to environmental perception. When the actual local situation conflicts with the assumptions of global planning, the system updates the semantic reliability of the relevant area and triggers replanning when necessary. This closed-loop process of "planning-execution-feedback-update" allows the robot to learn from failures and continuously optimize its environmental perception, thereby improving its adaptability and task success rate in complex dynamic environments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the semantic constraint path planning and dynamic obstacle avoidance method for quadruped robots provided by the present invention. Figure 2 This is a schematic diagram of the technical route of the semantic constraint path planning and dynamic obstacle avoidance method for quadruped robots provided by the present invention; Figure 3 This is a schematic diagram of global path planning in the semantic constraint path planning and dynamic obstacle avoidance method for quadruped robots provided by the present invention; Figure 4This is a schematic diagram of local control obstacle avoidance in the semantic constraint path planning and dynamic obstacle avoidance method for quadruped robots provided by the present invention; Figure 5 This is a schematic diagram of closed-loop update in the semantic constraint path planning and dynamic obstacle avoidance method for quadruped robots provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] Before providing a further detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0022] (1) Semantic reliability: refers to the degree to which a semantic label can be trusted under the current location, current time period, and current task conditions. Its value is calculated by combining semantic support with correction factors such as the robot's state and weather. Semantic reliability provides a decision basis for path planning that goes beyond purely geometric obstacles, enabling the robot to avoid areas that are geometrically passable but actually pose high risks or are unsuitable for passage.

[0023] (2) Semantic support: refers to the degree of support for a certain semantic category from multi-source information. It is obtained by weighted fusion of the output of the Visual Language Model (VLM), the output of the detection model, rule base constraints, and historical statistics. This support integrates high-level scene understanding, low-level object detection, prior rules, and long-term experience, providing a comprehensive and robust evidentiary basis for semantic judgment.

[0024] (3) Semantic blocking accumulation: This refers to the quantification of the inaccessibility risk at the semantic level in the A* algorithm. It is obtained by multiplying the inaccessibility judgments (0 or 1) of the four semantic categories (terrain, rules, dynamics, and task) with their respective semantic reliability and then summing the results. When this accumulation exceeds a preset threshold, the corresponding grid will be considered inaccessible, thus being effectively pruned in the path search, achieving the goal of transforming semantic risk into hard planning constraints.

[0025] (4) Semantic conflict index: This refers to a quantitative indicator used to judge whether the local execution situation is consistent with the global planning assumptions. It integrates situations such as local obstacle avoidance failure, local semantic risk exceeding expectations, path corridor deviation, and local stagnation. This index is the key to achieving global-local closed-loop control. When the index exceeds the threshold, it indicates that the semantic information on which global planning depends may have become invalid, and replanning or model update needs to be triggered.

[0026] (5) Path corridor: refers to a virtual channel defined around the global path skeleton, used to constrain the search range of the local obstacle avoidance module. Its width can be dynamically adjusted according to the safety margin and semantic information. The design of the path corridor gives the local controller the necessary degree of freedom while ensuring that its behavior does not deviate from the intention of the global plan. Especially when passing through narrow channels or densely populated areas, it can effectively prevent excessive avoidance actions from causing path instability.

[0027] like Figures 1 to 5 As shown, this application provides a semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots, aiming to solve the safety and efficiency problems of existing robot navigation methods in complex dynamic environments such as campuses, due to insufficient environmental representation, limited dynamic adaptability, and insufficient coordination between global planning and local execution. This method is executed by a computer device, specifically by a terminal or server alone, or by both. In this embodiment, the semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots includes the following steps S101 to S106. Wherein: Step S101: Obtain a multi-source observation set, which includes the target robot's state information, laser point cloud data collected by the target robot, and visual observation data. The target robot is a quadruped robot, such as a quadrupedal robot dog; application scenarios for quadruped robots include parks, campuses, warehouses, and logistics distribution.

[0028] Step S102: Assign a semantic label to each grid in the grid image of the scene where the target robot is located; calculate the initial confidence score of the semantic label based on the multi-source observation set; wherein the semantic label is used to characterize the semantic category of the semantic information of the scene where the target robot is located; the grid data is obtained by processing the laser point cloud data.

[0029] Step S103: Calculate the semantic reliability of the semantic tag based on the initial trust score. The semantic reliability is used to characterize the degree to which the semantic tag is trusted.

[0030] Step S104: Construct the cost function of the target robot based on the semantic reliability; based on the cost function, use a global path search algorithm to perform global path planning for the target robot to obtain a global path skeleton; the global path skeleton is used to guide the target robot from the task start point to the task end point.

[0031] Step S105: Generate the execution actions of the target robot during its movement along the global path skeleton based on the policy network. The execution actions include the linear velocity and angular velocity of the target robot.

[0032] Step S106: Determine whether the target robot's execution action was successfully executed. If the determination result is negative, update the semantic reliability and generate a global path skeleton and execution action based on the updated semantic reliability.

[0033] By implementing steps S101 to S106, this invention constructs a multi-dimensional semantic labeling system encompassing terrain, rules, dynamics, and tasks, quantifying it as semantic reliability and incorporating it into the cost function of global path planning. This enables the robot not only to recognize geometric obstacles but also to understand and adhere to the social attributes and traffic rules of the scene. This effectively prevents the robot from entering areas that are "geometrically passable but actually unsuitable for passage," making the planned path safer, more reasonable, and in line with social expectations. This invention employs a temporal reinforcement learning model capable of processing historical information for local obstacle avoidance. The robot can infer the movement trends of dynamic obstacles such as pedestrians and non-motorized vehicles based on continuous observation sequences. Compared to reactive obstacle avoidance relying solely on single-frame observations, this method can make more predictive decisions, effectively reducing hesitation, sudden stops, and frequent swaying in human-robot mixed scenarios, improving the stability and smoothness of local passage. By establishing a feedback update mechanism between global planning and local execution, this invention allows the system to transform "unexpected situations" encountered during local execution (such as path blocking or obstacle avoidance failure) into corrections to environmental perception. When local realities conflict with the assumptions of global planning, the system updates the semantic reliability of the relevant areas and triggers replanning when necessary. This closed-loop process of "planning-execution-feedback-update" enables the robot to learn from failures and continuously optimize its perception of the environment, thereby improving its adaptability and task success rate in complex and dynamic environments.

[0034] As one embodiment of this application, such as Figure 2As shown, in step S101, the laser point cloud data includes LiDAR data collected by sensors configured on the target robot. In this embodiment, the laser point cloud data is used for local geometric obstacle detection, grid occupancy estimation, and passable space analysis. The visual observation data includes RGB images collected by sensors configured on the target robot, specifically depth camera (RGB-D) images collected by a depth camera. In this embodiment, the visual observation data is used for target detection, semantic recognition, and high-level scene understanding. The target robot's state information includes odometry information, IMU / GNSS data, and body state bus data (body state bus data is used to characterize the target robot's body state information), used for pose estimation, velocity state estimation, slope judgment, battery reading, and joint state monitoring. After the above multi-source data are collected, they are input into the data layer as the data source for the target robot's path planning and dynamic obstacle avoidance.

[0035] Furthermore, such as Figure 1 The multi-source observation set also includes delivery task information provided by the task system, which includes the target robot's delivery start point, destination, delivery priority, payload mass, and current execution stage. In this embodiment, the multi-source observation set composed of the above data can be represented as: ; In the formula, Represents laser point cloud data; Represents visual observation data; and These represent odometer and inertial measurement information, respectively, where the inertial measurement information is IMU / GNSS data; Indicates the body's status information; This indicates delivery task information.

[0036] Furthermore, this application processes the laser point cloud data and visual observation data from the aforementioned multi-source observation set, unifying them into the robot's base coordinate system and the world coordinate system, so that the visual semantics can be subsequently projected onto the global or semi-global grid map; wherein, the formula for unifying the laser point cloud data and visual observation data from the multi-source observation set into the robot's base coordinate system and the world coordinate system is: ; In the formula, This represents the extrinsic transformation matrix from the sensor to the robot base; Indicates time The pose transformation matrix from the robot base to the world coordinate system. Estimated by fusion of odometry, IMU and positioning modules; Obtained from offline calibration; This represents multi-source observation data, including laser point cloud data and visual observation data. This application uses the above formula to ensure that laser, visual, and semantic information can be overlaid within the same map frame.

[0037] In step S102 of the embodiment of this application, as Figure 3 As shown in the multi-source semantic fusion module, the step of assigning semantic labels to each grid in the grid image of the scene where the target robot is located; calculating the initial confidence score of the semantic labels based on the multi-source observation set, includes the following steps S201 to S203. Wherein: Step S201: Assign a semantic label to each grid in the grid image according to the pre-constructed semantic labeling system. In this embodiment, the semantic labeling system includes at least four types of semantic labels: terrain semantics, rule semantics, dynamic semantics, and task semantics. Among them, terrain semantics is used to depict stairs, ramps, step edges, lawns, and unpaved roads; rule semantics is used to depict dormitory entrances, fire lanes, restricted areas on campus, and sidewalks; dynamic semantics is used to depict student traffic, shared bicycle accumulation, and construction barriers; and task semantics is used to depict express delivery stations, pickup points, and large / small dog fitting areas.

[0038] Step S202: The multi-source observation set is processed using the object detection model, the visual language model (VLM), and the rule base respectively to obtain the support of the object detection model for the semantic label, the support of the visual language model for the semantic label, and the support of the rule base constraints for the semantic label.

[0039] Step S203: The support of the target detection model for the semantic tag, the support of the visual language model for the semantic tag, the support of the rule base constraint for the semantic tag, and historical statistical data are fused to obtain the initial confidence score of the semantic tag.

[0040] As one embodiment of this application, step S202 involves processing the multi-source observation set using a target detection model, including: Target recognition is performed on the visual observation data in the multi-source observation set to output the target recognition result of the visual observation data in the current grid. The target recognition result is the target category, detection confidence and spatial location of at least one object. The current grid is any selected grid in the grid image.

[0041] Based on the semantic labels of the current grid, the support degree of the target recognition result for the semantic labels is calculated according to the target recognition result in the current grid based on the visual observation data; the specific steps are as follows: Define object category-semantic category association matrix In the formula, Indicating the target recognition result of the first The semantic category of each detected object. The semantic label representing the current grid The object category-semantic category association matrix is ​​used to characterize the contribution of the detected object to the semantic category in the target recognition result; for example, dynamic semantic tags. When representing "student-dense areas", the association weight of pedestrian objects is close to 1, while the association weight of objects such as fences and trash cans is low. The spatial influence of an object on the current grid in the target recognition result is calculated using a Gaussian kernel function; wherein, the formula for calculating the spatial influence of the current grid is: ; In the formula, This represents the distance from the projection center of the object onto the grid image to the current grid x. This indicates the range of influence in the control space, and its value ranges from 1.0 to 1.5 times the width of the target robot.

[0042] Based on the object category-semantic category association matrix and the spatial influence of the object on the current grid, the original evidence for the semantic label is calculated from the target recognition result output by the target detection model; this original evidence can be expressed as: ; In the formula, Indicating the target recognition results One object; This represents the detection confidence score of the i-th object in the target recognition result; Indicates the expansion of the projected area of ​​the object. The area of ​​influence obtained later; For the indicator function, when the grid If the value falls within this area, take 1; otherwise, take 0. The value of should be related to the target size and map resolution; a value of is recommended. A grid is used to mitigate edge errors of the detection frame.

[0043] The original evidence is subjected to exponential saturation normalization to obtain the support degree of the target recognition result for the semantic label; wherein, the formula for the exponential saturation normalization is: ; In the formula, This represents the support of the target recognition result output by the target detection model for the k-th semantic label in the grid x at time t; This is the scaling factor. It can be obtained by calibrating the "fit relationship between the original evidence and the semantic judgment accuracy" on the validation set.

[0044] In one embodiment of this application, step S202 involves processing the multi-source observation set using a Visual Language Model (VLM), including: The multi-source observation set is processed using a visual language model to obtain a structured description of the multi-source observation set; Based on the semantic tags of the current grid, a predefined set of semantic anchor points is defined. Calculate the matching degree between the structured description and the set of semantic anchors. ;in, This represents the j-th semantic anchor point of semantic tag k, for example, in dynamic semantic tags. When representing a "student-dense area", the set of semantic anchors can include "multiple pedestrians", "crossing behavior", "crowded central area", "need to give way", etc.

[0045] The original evidence of the visual language model for the semantic label of the current grid is calculated based on the matching degree between the structured description and the set of semantic anchors; wherein, the original evidence of the visual language model for the semantic label of the current grid can be represented as: ; In the formula, , Indicates the first The importance weight of each semantic anchor point; This represents the degree of matching between the set of semantic anchors of the structured description and the semantic tag k.

[0046] The original evidence of the visual language model for the semantic labels of the current grid is normalized to obtain the support level of the visual language model for the semantic labels of the current grid; wherein, the normalization process uses the Sigmoid function, specifically expressed as: ; In the formula, In this embodiment, ; This represents the stretching factor, used to control the sensitivity of the original VLM evidence to support. The neutral threshold represents the "boundary point between support and non-support". and It can be obtained from the offline validation set by minimizing the binary cross-entropy.

[0047] In one embodiment of this application, step S202 involves processing the multi-source observation set using a rule base to obtain the support degree of the rule base constraints for semantic tags, including: Set a set of association rules for the semantic labels of the current grid, the set of association rules including at least one rule; the rule is one of campus traffic rules, social norms, task constraints, and time period characteristics; the set of association rules can be represented as: ; In the formula, This represents the j-th rule in the set of association rules.

[0048] Calculate the satisfaction degree of each rule in the association rule set, and set the rule weight of the rule; The original evidence for the semantic tag is calculated by the rule base based on the satisfaction level and the rule weight; wherein the calculation formula for the original evidence for the semantic tag by the rule base is: ; In the formula, Representation rules satisfaction level Representation rules The rule weights.

[0049] The original evidence of the semantic tag from the rule base is used as the support level of the rule base for the semantic tag; the support level of the rule base for the semantic tag can be expressed as: ; It should be noted that satisfaction level The calculation method depends on the rule type. For example, if the rule is "Increased pedestrian risk at the intersection towards the cafeteria from 11:30 AM to 12:30 PM", then if the current time falls within that time window and the grid... The satisfaction level is close to 1 when the location is near the intersection; if it is outside the specified time period or location, the satisfaction level is lower. Rule weight. The importance of the rule should be determined by its weight, and it is recommended to set it based on a combination of "the severity of the consequences of the violation and the impact on path security".

[0050] In one implementation of this application, in step S203, the historical statistical data is used to describe the frequency with which the semantic category of the current grid has been represented by semantic label k under similar time periods and similar tasks. To avoid outdated data interfering with the current judgment, a weighted statistical method with time decay is used. Let there be H observations matching the current conditions within the historical window, and the... The observation occurred at time 10:00. The corresponding semantic verification result is Therefore, original historical evidence is defined as: ; in, Indicates the time decay coefficient. , The closer it is to 1, the more the system values ​​long-term stable statistics; The smaller the value, the more the system prioritizes recent observations. For dynamic semantics related to campus pedestrian flow, it is recommended to take... For long-term stable static task semantics (such as express delivery stations, dormitory entrances), a value closer to 1 can be taken.

[0051] Using the aforementioned original historical evidence as the support level of the historical statistical data for the semantic label k, it can be specifically expressed as follows: .

[0052] As one embodiment of this application, in step S203, the calculation formula for fusing the support of the target detection model for the semantic tag, the support of the visual language model for the semantic tag, the support of the rule base constraints for the semantic tag, and historical statistical data is as follows: ; In the formula, These represent the degree of support for semantic label k by the output of the Visual Language Model (VLM), the output of the detection model, the rule constraints, and historical statistical data, respectively. To normalize the fusion weights, satisfy . and Accuracy evaluation from offline semantic validation sets, Given the importance of the manually assigned rules, This indicates the impact of historical experience on current identification.

[0053] It should be noted that the result calculated from the above formula... Let k be the initial confidence score of the semantic label k for each grid in the grid image, representing the degree to which semantic label k supports the passage of the current grid x. Since the semantic label system pre-constructed in this application includes at least four types of semantic labels: terrain semantics, rule semantics, dynamic semantics, and task semantics, therefore, semantic labels... ,in, Representing terrain semantics, Representing the semantics of rules, Represents dynamic semantics. The task semantics are represented as follows: For each grid, the support of that grid in four aspects—terrain semantics, rule semantics, dynamic semantics, and task semantics—is calculated, i.e., the initial confidence score. Since the main evidence sources for different semantic labels differ, terrain semantics mainly relies on laser point clouds, visual observation, terrain recognition results, and historical traffic statistics; rule semantics mainly relies on the campus rule base, prior maps, current location relationships, and time period conditions, with the detection model and VLM output serving as supplementary evidence; dynamic semantics mainly relies on the detection model, VLM structured output, and historical dynamic statistics; and task semantics is mainly determined by the task system input, task area configuration, and the relationship between the current location and the task objective. For evidence sources that are not applicable in a certain semantic label, their corresponding support can be set to 0, or they can be excluded from the fusion process.

[0054] As one embodiment of this application, such as Figure 3 As shown in the semantic reliability model, in step S103, the semantic reliability of the semantic tag calculated based on the initial confidence score is output by the semantic reliability model, specifically including the following steps S301 to S302. Wherein: Step S301: Obtain the correction factor of the semantic reliability, the correction factor including the correction factor determined by the physical state of the target robot and the weather condition correction factor; Step S302: Based on the initial confidence score and the correction factor, calculate the semantic reliability of the semantic tag using a truncation function; wherein, the formula for calculating the semantic reliability of the semantic tag is: ; In the formula, Indicates the initial confidence score; This represents a correction factor determined by the organism's state, where... It can include payload mass, electrical charge, and attitude information; This represents the weather condition correction factor, where, It can indicate external environmental conditions such as sunny day, cloudy day, rainy day, water accumulation, and slippery surface; These represent the upper and lower limits of semantic reliability, respectively. This represents the truncation function.

[0055] It should be noted that the semantic reliability is used to describe the degree to which the semantic tag can be trusted under the current location, current time period, and current task conditions; the meaning of the semantic reliability calculation formula is that even if the same location and the same semantic tag are usually accessible, if the weather deteriorates or the machine is in poor condition, its reliability will decrease, thereby promoting a more conservative global planning.

[0056] As one embodiment of this application, step S104, constructing the cost function of the target robot based on the semantic reliability, includes the following steps S401 to S404. Wherein: Step S401: Assign binary scores to each semantic tag in the semantic tagging system. In this embodiment, the semantic tagging system mainly includes terrain semantics, rule semantics, dynamic semantics, and task semantics, and the binary scores corresponding to the above four types of semantics are denoted as follows: , , and The binary scores of the above four semantic categories can be represented as: ; In the formula, This represents the semantic determination result of the lattice x for the u-th class at time t; This indicates the permissible semantic library within this type of semantics; This refers to the inaccessible semantic library within this semantic category. In this application, accessible semantic libraries are uniformly categorized as follows: Unusable semantic libraries are uniformly categorized into Therefore, the score itself does not express the magnitude of the risk, but only whether "this type of semantics considers the current area as a no-entry zone".

[0057] Step S402: Calculate the semantic blocking accumulation of the current grid based on the semantic reliability and the binary score of the semantic label. The semantic blocking accumulation is the sum of the products of the binary score of each semantic label and the corresponding semantic reliability. The semantic blocking accumulation is used to characterize the comprehensive strength of the impassable judgment of the current grid. The current grid is the selected grid in the grid image.

[0058] In this embodiment of the application, the reliability of the above four types of semantics at the current location is defined as follows: , , and The semantic blocking accumulation can then be expressed as: ; In the formula, This represents the combined strength of the four semantic categories in determining whether the current grid is "unpassable". Since the score for the passable semantic category is... Therefore, even if a certain type of semantic has high reliability, it will not increase the cumulative blocking amount; only when a certain type of semantic is classified into the inaccessible semantic library will its corresponding reliability be added in a weighted manner. When the sum of "reliability × score" of the four types of semantic is greater than 1, the grid is directly regarded as inaccessible.

[0059] Step S403: Determine whether the current grid is passable based on the accumulated semantic blocking amount, and determine the semantic reliability cost based on the determination result; wherein, determining whether the current grid is passable based on the accumulated semantic blocking amount includes: When the accumulated semantic blocking value is greater than 1, the current grid is determined to be impassable; when the accumulated semantic blocking value is less than or equal to 1, the current grid is determined to be passable; the determination process can be expressed as follows: ; In the formula, This indicates that grid x is directly determined to be impassable at time t, and the node will not be expanded further. This indicates that the grid is still allowed to enter the candidate search set. A threshold of 1 means that the system will only truly block the grid when at least one of the four semantic categories is reliably determined as prohibited, or when multiple categories of moderately reliable prohibition judgments are superimposed to a sufficiently high level. This avoids overly conservative decisions based on single weak evidence, while also preventing obvious multi-source prohibition information from being ignored.

[0060] The semantic reliability cost is determined based on the judgment result, including: when the judgment result is that the current grid is impassable, the semantic reliability cost is determined to be infinite; when the judgment result is that the current grid is passable, the accumulated semantic blockage amount is used as the semantic reliability cost; wherein, the method for determining the semantic reliability cost can be expressed as: ; Based on the above method for determining semantic reliability cost, this application can use the accumulated blocking amount itself as the residual semantic reliability cost for nodes that are still allowed to be searched; if a node has been determined to be impassable, then an infinite penalty is directly imposed.

[0061] Step S404: Construct the cost function of the target robot based on the semantic reliability cost; the cost function of the target robot includes geometric collision risk, semantic reliability cost, and quadruped robot-specific motion cost, and the expression of the cost function is: ; In the formula, This represents the cost of geometric occupancy and barrier expansion; This represents the cost associated with semantic reliability; This represents the cost of quadrupedal locomotion. Weighting. The principle of safety priority must be met under passable conditions: generally speaking The largest, followed by And then The final weight values ​​were optimized through offline replay experiments, and the objective function was a weighted result of "passability, maximum collision-free rate, and minimum average latency".

[0062] In this embodiment, the movement cost specific to the quadruped robot is composed of factors such as slope, steps, slipperiness, load, and power consumption. Therefore, the movement cost specific to the quadruped robot can be expressed as: ; In the formula, Indicates slope; Indicates the step / drop indicator function; Indicates the risk of slippery conditions; Indicates the current quality of the goods. This is the nominal reference load, and SOC indicates the battery's state of charge. The above parameters are sourced from: slope. Estimated from topographic layers or height differences; Obtained by step edge recognition; Data can be derived from visual ground material classification or historical slip event statistics; load mass and SOC are directly derived from the task system and robot state bus.

[0063] It should be noted that, Indicates the slope cost weight, Indicates the weight of the step / drop cost. Indicates the weight of the risk cost of slippery conditions. Indicates load cost weight, This indicates the weight of the electricity cost. , , , as well as All are non-negative coefficients, used to adjust the relative importance of each motion-influencing factor in the quadruped robot's specific motion cost.

[0064] Specifically, Used to adjust the degree of influence of slope on the cost of movement. The larger the value, the more sensitive the system is to changes in slope, and the more likely the planning will avoid areas with large slopes. Used to adjust the degree of influence of steps or elevation differences on the cost of movement. The larger the area, the more the system tends to avoid areas with steps, drops, or edge risks; Used to adjust the degree of impact of slippery risk on the cost of sports. The larger the value, the more the system tends to avoid slippery surfaces; Used to adjust the degree of influence of load mass on motion cost. The larger the size, the higher the requirements for path stability and safety of the robot under heavy load; Used to adjust the degree to which the battery's state of charge affects the cost of movement. The larger the value, the more likely the system is to choose a path with lower energy consumption and more stable passage when the battery is low. The weights mentioned above need to be set according to the quadruped robot model, body width, load capacity, underlying gait control performance, and experimental results in typical campus scenarios, so as to achieve a balance between safety, stability, and passage efficiency in the path planning results.

[0065] As one embodiment of this application, such as Figure 3 As shown in the semantic reliability-based A* global path planning, in step S104, based on the cost function, a global path search algorithm is used to perform global path planning for the target robot to obtain the global path skeleton, including the following steps S501 to S502. Wherein: Step S501: Obtain the path predecessor node obtained in the previous search stage, take the path predecessor node as the current node, and determine the path predecessor node of the current node in the current search stage according to the global path search algorithm; wherein, the path predecessor node is the adjacent node that directly generates the current node. It can be understood that the current node used by the global path search algorithm in the initial stage of executing the path search is the delivery starting point.

[0066] The step of determining the predecessor node of the current node in the current search phase according to the global path search algorithm includes the following steps S601 to S603. Wherein: Step S601: Perform a passage determination on at least one target node. When the target node meets the passage determination requirements, add the target node as a candidate node to the open list. The target node is the adjacent node of the current node. In this embodiment, each grid in the grid image of the scene where the target robot is located is considered as a node. The passage determination requirements are that the grid corresponding to the current node is passable and the geometric occupancy and obstacle expansion cost of the grid is less than a preset geometric obstacle cost threshold. The passage determination requirements can be expressed as follows: ; In the formula, This represents the geometric barrier cost threshold. The formula indicates that a node is allowed to enter the open list only if semantically it is not blocked and its geometric barrier cost does not exceed the safety threshold.

[0067] Step S602: For all candidate nodes in the open list, calculate the travel cost of all candidate nodes using an evaluation function; wherein the expression of the evaluation function is: ; In the formula: Represents the evaluation function; The inheritance cost function represents the cost from the current node. iWalk to the predecessor node of the path j The cost, Indicates traversing to the current node. i Costs already incurred; This is a heuristic function used to compute the current node. i With the path's predecessor node j The cost of traveling the straight-line distance (Euclidean distance) between them; Indicates starting from the current node i To the predecessor node of the path j The geometric distance. Considering that frequent large-angle turns of a quadruped robot under heavy load will increase posture sway and energy consumption, the cornering penalty coefficient is increased. To avoid excessive polyline rendering of the global skeleton; This represents a two-step heading angle change. The formula uses Euclidean distance as a heuristic.

[0068] Step S603: Select the candidate node with the lowest travel cost as the predecessor node of the path obtained in the current search stage.

[0069] Step S502: Determine whether the predecessor node of the path obtained in the current search stage is the delivery destination. If the result is no, repeat step S501; otherwise, generate a global path skeleton based on the delivery origin, delivery destination, and predecessor nodes of the path obtained in all search stages. P ={ x 0,…, x N Furthermore, it should be noted that if the open list is empty and the delivery destination is still not found, it is determined that there is no feasible path under the current semantic reliability cost map, and a replanning or reduction of constraints is triggered to search again.

[0070] Please refer to it again. Figure 3 , Figure 3 The semantic reliability-based A* global path planning is used to represent steps S103 and S104 above. Specifically, Figure 3The grid area in the map is divided into a regular two-dimensional grid map, representing the robot's global exploration space. The grids in the map are presented in two states: white and black. White represents passable areas, and black represents obstacle areas. Among them, obstacle areas are static physical obstacles (such as walls and fixed facilities) that are identified by sensors as absolutely impassable. In planning, they are usually set as mandatory no-entry nodes. The grids scattered around the obstacle areas in the map represent semantic blocking grids. Semantic blocking grids are not absolute physical obstacles, but areas marked as "untrustworthy" or "high-risk". For example, there are semantically ambiguous areas with low recognition confidence due to insufficient light or areas where execution has failed in the past. This application uses the A* algorithm to actively avoid these grids as high-cost areas during route planning, thereby providing greater safety redundancy than traditional geometric obstacle avoidance.

[0071] As one embodiment of this application, step S105, which involves generating the target robot's actions along the global path skeleton based on the policy network, includes steps S701 to S703. Wherein: Step S701: During the movement of the target robot along the global path skeleton, acquire the local observation vector and historical information set of the target robot at the current moment, and extract the short-term historical features of the historical information set. The local observation vector can be represented as: ; In the formula, Represents the local observation vector; This represents the lidar scanning sequence; and These represent the polar radius and azimuth angle of the current sub-target point relative to the robot, respectively. The current sub-target point is a target node in a local execution stage of the target robot's movement along the global path skeleton. The movement of the target robot along the global path skeleton is deconstructed into several local execution stages. The current sub-target of each stage corresponds to the expected pose node at the next moment. By planning the sub-targets of each local node step by step, the robot's entire execution action is generated. The local semantic risk vector can be composed of factors such as pedestrian flow risk, entrance congestion, and construction site enclosure risk. This indicates the lateral deviation of the robot relative to the path corridor; This indicates the action performed at the previous moment, used to characterize the robot's speed inertia, turning trend, and short-term obstacle avoidance continuity.

[0072] It should be noted that, in this embodiment, the path corridor is used to provide local obstacle avoidance space for the target robot to travel along the global path skeleton. In this embodiment, the path corridor is generated based on the global path skeleton and specifically includes: Using the global path skeleton as the centerline of the path corridor, the corridor half-width is calculated based on the target robot's body width and lateral safety margin. The path corridor is generated based on the centerline and the corridor half-width; wherein, the path corridor can be represented as: ; In the formula, Indicates half the width of the corridor. ,in, Indicates the width of the fuselage. This indicates lateral safety margin, specifically in narrow corridors and doorways, where half the corridor width is considered safe. It should also be multiplied by a semantic contraction factor. This helps prevent excessive deviation of local control in narrow passages and areas where crowds gather. Represents the global path skeleton.

[0073] It should be noted that, as Figure 4 As shown, Figure 4 The path corridor and environment model in the model represent the global path skeleton of the target robot. Figure 4 The planned path is a curved path (and path corridors are dashed lines distributed on both sides of the global path skeleton); by Figure 4 As can be seen, the global path skeleton is restricted to the area of ​​the path corridor to ensure that the target robot (robot dog) maintains a safe distance when it approaches obstacles (school fences, bicycles, and pedestrians, etc.), thereby planning a driving path for the target robot that can avoid obstacles.

[0074] Furthermore, the local semantic risk vector in this embodiment is defined as: ; In the formula, This indicates semantic risks related to terrain, primarily reflecting the impact of local terrain conditions such as steps, ramps, slippery surfaces, and narrow passages on traffic safety. This indicates semantic risks in the rules, mainly reflecting constraints such as restricted areas, separation of pedestrians and vehicles, time-limited passage, and right-of-way rules. This indicates dynamic semantic risks, mainly reflecting time-varying risks such as pedestrian flow, shared bicycle accumulation, temporary barriers, and mobile obstacles; This indicates semantic risk of the task, mainly reflecting task conditions such as delivery priority, load sensitivity, and arrival operation stage.

[0075] It should also be noted that the historical information set includes local observation vectors at several consecutive moments, the target robot's execution actions, and reward values ​​calculated by the local reward function. The historical information set is used to describe the motion trend of dynamic obstacles, the robot's short-term obstacle avoidance behavior, and the local conflict change process. Since obstacles such as pedestrians, bicycles, and shared bicycle piles in a campus scene have temporal correlations, historical information can compensate for the difficulty of fully describing the state of dynamic obstacles with a single frame observation. In step S701 of this embodiment, short-term historical features of the historical information set are extracted to refine the motion trend of dynamic obstacles, repetitive obstacle avoidance behavior, and short-term conflict patterns. Specifically, this embodiment uses LSTM (Local Temporal Reinforcement Learning Control Module) to extract the short-term historical features of the historical information set. These short-term historical features can be expressed as: ; In the formula, Indicates the length of the history window.

[0076] Furthermore, in this embodiment of the application, the formula for calculating the reward value is: ; In the formula, Indicates the robot's current state. Distance to the sub-target point This indicates that the robot was at the previous moment. Distance to the sub-target point; Indicates the robot's current state. linear velocity; Indicates the robot's current state. angular velocity, This indicates that the robot was at the previous moment. angular velocity; To reach the threshold, a length of 0.4-0.6 times the robot's length is typically used. This is the minimum detection distance for lasers. The collision threshold is typically expressed as "half the width of the aircraft's outer profile + sensor error + minimum safety margin"; For the lateral deviation of the corridor, For local semantic risks; Indicates reaching a positive reward; Indicates a collision penalty; Used to control the importance of moving towards the goal Used to encourage increased linear speed while ensuring safety. Used to suppress frequent head shaking; Used to maintain the consistency of the path corridor This is used to penalize violations of social semantics and high-risk semantic areas. The calculation formula for this reward function can be simply expressed as: reward for getting closer to the target + reward for moving quickly - penalty for frequent turns - penalty for running out of the corridor - penalty for approaching high-risk areas.

[0077] Step S702: A policy network is used to process the local observation vector and short-term historical features at the current moment to obtain the basic actions of the target robot; an expert experience network is used to process the local observation vector and local semantic risk vector at the current moment to obtain the auxiliary actions of the target robot. This step S702 is specifically performed by… Figure 4 The TS-TD3 local controller of the target robot executes the command. Figure 4 In this context, the Actor network represents the policy network, and the dual Critic network represents the dual evaluation network used during the training of the policy network.

[0078] In step S702, the basic action can be represented as: ; In the formula, Representational Policy Network The basic actions of output; Represents the local observation vector; Indicates short-term historical characteristics; Indicates the intensity of the noise during exploration.

[0079] In step S702, the auxiliary action can be represented as: ; In the formula, Indicating expert experience network The auxiliary actions output; Represents the local observation vector; This indicates local semantic risk.

[0080] Step S703: Determine the basic action or auxiliary action with the highest action value as the action to be executed by the target robot, which can be specifically expressed as: ; In the formula, This represents the action evaluation function, used to assess the action value of the basic action and the auxiliary action. It should be noted that the executed action is given by the target robot's velocity layer, and the executed action can be represented as: ; In the formula, Indicates linear velocity. This represents angular velocity. For quadruped robot dogs, the upper-level navigation typically only outputs velocity-level commands, while the specific gait, foot trajectory, and body stability control are handled by the lower-level motion controller. Therefore, it is unnecessary to directly incorporate joint angles into the navigation motion space.

[0081] It should be noted that, as can be seen from the expression in step S702 above, this application combines the local semantic risk vector. Constructed semantic expert actions. Because Composed of terrain semantic risk, rule semantic risk, dynamic semantic risk, and task semantic risk, expert actions can respond not only to dynamic risks such as pedestrian flow, shared bicycle accumulation, and barriers, but also to terrain risks such as step edges and slippery surfaces, rule risks such as no-entry / yield restrictions, and task risks such as vulnerable loads or high-time-sensitive tasks. For example: when terrain semantics indicate the presence of step edges or slippery areas ahead, linear velocity is reduced; when rule semantics indicate a prohibition on crossing, a detour suggestion is directly given; when the risk of pedestrian flow or shared bicycle accumulation is high in dynamic semantics, lateral avoidance offset is increased; when task semantics indicate that the current load is vulnerable or delivery priority is high, the head-swing amplitude is appropriately tightened and a smoother control output is maintained.

[0082] Traditional navigation processes typically follow a "plan first, execute later" approach, meaning a global path is first generated, and then executed by a local planning or control module. This process can accomplish basic navigation tasks in relatively stable environments, but in semi-open campus environments, the local execution phase may encounter situations not fully anticipated in the global planning phase, such as temporary obstacles, blocked passageways, changes in pedestrian density, or the robot repeatedly avoiding obstacles in local areas. This invention utilizes a global-local closed-loop update mechanism, allowing the results of local execution to influence subsequent planning decisions. Specifically, in step S106 of this embodiment, determining whether the target robot's action was successfully executed, and updating the semantic reliability when the result is negative, includes steps S801 to S802. Wherein: Step S801: Calculate the semantic conflict index based on the execution result of the action and the semantic risk vector; wherein, the formula for calculating the semantic conflict index is: ; In the formula, This indicates that several consecutive steps in a localized area cannot generate an executable safe action; This represents the local semantic risk in global prediction; This represents the local semantic risk obtained during the local execution phase based on actual observations; Indicates local stagnation / repeated detour; parameters Recommended to meet Maximum. The execution results include case 1: several consecutive steps in a local area fail to generate an executable safe action and case 2: a local area gets stuck / repeatedly detours. The fact that the action in case 1 cannot be executed indicates that the global semantics have been significantly distorted.

[0083] Step S802: Determine whether the semantic conflict index is greater than the trigger threshold. If the determination result is yes, update the semantic reliability. The expression for updating the semantic reliability is: ; This formula represents the reliability update method, where, It is the local reliability inferred from actual observations and action feasibility during the local execution phase; This represents the semantic reliability output by the semantic reliability model during the local execution phase; For the update rate, a value of 0.1–0.3 is recommended. If long-term statistical stability is emphasized later, a Beta-Bernoulli update can also be used: success is recorded as 1, and failure or strong conflict is recorded as 0, thus accumulating the posterior distribution of semantic reliability. It should be noted that the trigger threshold is determined through repeated experimental debugging. In a typical campus scenario, the distribution of successful execution samples and failed execution samples is statistically analyzed, and the boundary between the two or the optimal point of F1 (the harmonic mean of precision and recall) is taken as the initial value of the threshold, and then fine-tuned through real machine or simulation playback.

[0084] It is understandable that step S105 is a local execution phase during the global movement of the target robot (along the global path skeleton), and after the local execution phase is completed, a local execution result is output. For example... Figure 5 As shown, Figure 5 The system demonstrates a semantic conflict detection and global-local closed-loop update mechanism, specifically the method in step S106 for updating semantic reliability based on execution results. Specifically, the system first performs semantic consistency analysis on continuously received local sensor data, movement trajectories, and semantic recognition results. Then, it calculates a semantic conflict index based on the globally predicted local semantic risk and the local semantic risk obtained from actual observations during the local execution phase, and performs a threshold determination. If the determination is negative (No), meaning the conflict has not exceeded the threshold, the system is deemed trustworthy and continues local execution to maintain the current state. If the determination is positive (Yes), meaning the conflict exceeds the threshold, semantic conflict processing is triggered, and a global-local closed-loop correction is performed: on one hand, the semantic reliability is updated and fed back to the A* global path planning module to adjust the current local motion; on the other hand, the global semantic reliability is corrected and fed back to the semantic reliability model. Subsequently, global replanning is triggered to update the global path strategy, and the new global information is input again into the conflict determination stage, thus forming a complete global-local closed loop.

[0085] because Based on the target robot's actual perception of the environment and the feasibility of its actions, the target robot did not output any action at the delivery starting point. =0 / undefined Calculated by the semantic reliability model, When the target robot moves from the delivery starting point to the next sub-target node, the target robot calculates semantic reliability based on actual observations and action feasibility during this local execution phase. and use and delivery origin The semantic reliability is updated to obtain a new semantic reliability, i.e. .

[0086] By implementing step S106 above, the present invention can transform the navigation process from a unidirectional "global planning - local execution" to a closed-loop "global planning - local execution - feedback judgment - reliability update - replanning when necessary". This closed-loop mechanism can reduce the occurrence of repeated detour failures or short-term pauses of the robot in the same area, and improve the system's adaptive adjustment capability in the face of dynamic environmental changes.

[0087] Furthermore, this application also provides a training process for the policy network, which is used to update the policy network and evaluate its parameters, enabling the policy network to output safer, smoother, and semantically constrained actions in dynamic obstacle environments. The specific training process includes steps one through eight. Wherein: Step 1: Initialize the policy network Evaluation Network Target-Policy Network and target evaluation network The system initializes the experience replay pool and the historical experience pool. The policy network generates continuous execution actions, the evaluation network assesses the value of a combination of observed states and actions, the target network constructs a stable training objective, the experience replay pool stores single-step samples, and the historical experience pool stores continuous sets of historical information.

[0088] Step 2: At each time t, the system calculates the current local observation vector. Generate execution actions from historical information sets The robot then sends the action to its underlying controller. After the robot completes the action, the environment returns to the observation state for the next moment. And calculate the reward value based on the local reward function. The reward value takes into account factors such as proximity to the target, collision penalty, speed bonus, smooth steering, path corridor deviation, and local semantic risk.

[0089] Step 3: Convert the local observation vector Historical information collection Execution of actions Reward Value and the local observation vector at the next time step The samples are stored in the experience replay pool. During training, N samples are randomly sampled in batches from the experience replay pool, and the corresponding historical information set is read from the historical experience pool. These samples are then input into the LSTM module for feature extraction to obtain the motion trend information of the dynamic obstacles.

[0090] Step 4: Utilize the target policy network Generate the target action for the next moment. and target action Add smooth noise This process yields the smoothed target action. This step helps prevent drastic jumps in the policy network's output action and reduces the evaluation network's overestimation of local anomalous actions. The smoothed target action can be represented as: ; Step 5: Utilize two target evaluation networks and The two values ​​are applied to the smoothed target action respectively, and the smaller value is used to construct the temporal difference target. This step is used to alleviate the overestimation problem of the evaluation network and improve training stability. The constructed temporal difference target can be expressed as: ; Step Six: Based on the time series difference objective Compared with the current evaluation network output The error between them is used to construct an evaluation network loss. The evaluation network parameters are updated by minimizing this loss. The purpose of the evaluation network loss is to enable the evaluation network to more accurately estimate the value of the current observation and action combination. Specifically, the evaluation network loss... It can be represented as: ; Step 7: Update the policy network according to the delayed update mechanism. Policy network updates are achieved through policy gradients, which adjust the policy network parameters based on the action values ​​given by the evaluation network, so that the policy network outputs actions that yield higher expected returns. The expression for the policy gradient is: ; In the formula, , representing the expected discounted return of strategy learning; The discount rate is denoted as 0.98-0.995, preferably 0.99. This value indicates that the model focuses on both short-term obstacle avoidance safety and retains a sufficiently long target-reaching drive. To provide a reference quantity related to the complete historical trajectory for the policy gradient, thereby reducing update variance and enhancing the model's sensitivity to hidden dynamic states, this application also sets a discount reward baseline during training. The expression for the historical baseline is: .

[0091] Step 8: Perform soft updates on the target policy network and the target evaluation network. Update the target network parameters slowly at a certain ratio to allow the target network to smoothly follow the current network, thereby improving the stability of the training process. The soft update process for the target policy network and the target evaluation network can be represented as follows: ; In the formula, Represents the target evaluation network Network parameters, Indicates evaluation network Network parameters; Represents the target policy network Network parameters, Representational Policy Network Network parameters.

[0092] In summary, by implementing steps S101 to S106, this invention, through semantic environment representation, semantic reliability modeling, global semantic path planning, local temporal reinforcement learning control, and closed-loop feedback updates, enables the quadruped robot to possess stronger passage judgment capabilities, dynamic obstacle avoidance capabilities, and adaptive adjustment capabilities in complex semi-open scenarios such as campus logistics. These effects are primarily based on theoretical analysis of the system structure and algorithm mechanism, and can be further quantified and verified through simulation experiments or real-world testing, such as comparing indicators like path success rate, collision rate, average passage time, repeated obstacle avoidance counts, and task completion rate under different methods.

[0093] In one exemplary embodiment, this application also provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0094] In one exemplary embodiment, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0095] In one exemplary embodiment, this application provides a computer program product including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0096] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0097] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0098] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A semantically constrained path planning and dynamic obstacle avoidance method for a quadruped robot, characterized in that, include: Acquire a multi-source observation set, which includes the target robot's state information, laser point cloud data collected by the target robot, and visual observation data; Assign a semantic label to each grid in the grid image of the scene in which the target robot is located; calculate an initial confidence score for the semantic label based on the multi-source observation set; wherein the semantic label is used to characterize the semantic category of the semantic information of the scene in which the target robot is located; The semantic reliability of the semantic tag is calculated based on the initial confidence score, and the semantic reliability is used to characterize the degree to which the semantic tag is trusted. A cost function for the target robot is constructed based on the semantic reliability. Based on the cost function, a global path search algorithm is used to perform global path planning for the target robot to obtain a global path skeleton. The global path skeleton is used to guide the target robot from the task start point to the task end point. The action to be performed by the target robot during its movement along the global path skeleton is generated based on the policy network; Determine whether the target robot's action was successfully executed. If the determination result is negative, update the semantic reliability and generate a global path skeleton and action based on the updated semantic reliability.

2. The semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots according to claim 1, characterized in that, The process involves assigning semantic labels to each grid in the grid image of the scene in which the target robot is located; The initial confidence score of the semantic label is calculated based on the multi-source observation set, including: Each grid in the grid image is assigned a semantic label according to a pre-constructed semantic labeling system; the semantic labeling system includes at least terrain semantics, rule semantics, dynamic semantics, and task semantics. The multi-source observation set is processed using an object detection model, a visual language model, and a rule base to obtain the support of the object detection model for the semantic label, the support of the visual language model for the semantic label, and the support of the rule base constraints for the semantic label. The initial confidence score of the semantic tag is obtained by fusing the support of the target detection model for the semantic tag, the support of the visual language model for the semantic tag, the support of the rule base constraint for the semantic tag, and historical statistical data.

3. The semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots according to claim 1, characterized in that, The step of calculating the semantic reliability of the semantic tag based on the initial confidence score includes: Obtain the correction factor for the semantic reliability, which includes a correction factor determined by the physical state of the target robot and a weather condition correction factor; The semantic reliability of the semantic tag is calculated using a truncation function based on the initial confidence score and the correction factor.

4. The semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots according to claim 2, characterized in that, The step of constructing the cost function for the target robot based on the semantic reliability includes: Assign a binary score to each semantic tag in the semantic tagging system; The semantic blocking accumulation is calculated based on the semantic reliability and the binary score of the semantic label. The semantic blocking accumulation is the sum of the products of the binary score of each semantic label and the corresponding semantic reliability. The current grid is the selected grid in the grid image. Based on the accumulated semantic blocking amount, determine whether the current grid is passable, and determine the semantic reliability cost based on the determination result; Construct the cost function of the target robot based on the semantic reliability cost; The determination of whether the current grid is passable based on the accumulated semantic blocking amount includes... When the accumulated semantic blocking amount is greater than 1, the current grid is determined to be impassable; when the accumulated semantic blocking amount is less than or equal to 1, the current grid is determined to be passable. The semantic reliability cost is determined based on the judgment result, including: When the judgment result is that the current grid is impassable, the semantic reliability cost is determined to be infinite; when the judgment result is that the current grid is passable, the semantic blocking accumulation is used as the semantic reliability cost.

5. The semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots according to claim 1, characterized in that, The cost function includes geometric collision risk, semantic reliability cost, and quadruped robot-specific motion cost.

6. The semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots according to claim 1, characterized in that, The step of using a global path search algorithm to plan a global path for the target robot based on the cost function to obtain a global path skeleton includes: Obtain the path predecessor node obtained in the previous search stage, use the path predecessor node as the current node, and determine the path predecessor node of the current node in the current search stage according to the global path search algorithm; wherein, the path predecessor node is the adjacent node that directly generates the current node, and the current node used by the global path search algorithm in the initial stage of executing the path search is the delivery starting point. Determine whether the predecessor node of the path obtained in the current search stage is the delivery destination. If the result is no, repeat the step of obtaining the predecessor node of the path obtained in the previous search stage, taking the predecessor node of the path as the current node, and determining the predecessor node of the current node in the current search stage according to the global path search algorithm; otherwise, generate a global path skeleton based on the delivery origin, delivery destination, and the predecessor nodes of the path obtained in all search stages.

7. The semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots according to claim 6, characterized in that, The step of determining the predecessor node of the current node in the current search phase according to the global path search algorithm includes: A passage determination is performed on at least one target node. When the target node meets the passage determination requirements, the target node is added to the open list as a candidate node. The target node is an adjacent node that is adjacent to the current node. For all candidate nodes in the open list, the evaluation function is used to calculate the travel cost of all candidate nodes; The candidate node with the lowest travel cost is selected as the predecessor node of the path obtained in the current search stage.

8. The semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots according to claim 1, characterized in that, The actions performed by the target robot during its movement along the global path skeleton, generated based on the policy network, include: During the movement of the target robot along the global path skeleton, the local observation vector and historical information set of the target robot at the current moment are obtained, and the short-term historical features of the historical information set are extracted. A policy network is used to process the local observation vector and short-term historical features at the current moment to obtain the basic actions of the target robot; an expert experience network is used to process the local observation vector and local semantic risk vector at the current moment to obtain the auxiliary actions of the target robot. The basic or auxiliary actions with the greatest value are identified as the actions to be executed by the target robot. The local observation vector includes a lidar scanning sequence, the polar radius and azimuth of the current sub-target point relative to the robot, and the current sub-target point is a target node in a local execution phase of the target robot's movement along the global path skeleton, a local semantic risk vector, the lateral deviation of the target robot relative to the path corridor, and the action performed by the target robot at the previous moment; the path corridor is generated based on the global path skeleton; the historical information set includes multiple historical information items arranged in chronological order, and the historical information includes observation data, the target robot's execution actions, and the reward value calculated by the local reward function.

9. The semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots according to claim 8, characterized in that, The path corridor is generated based on the global path skeleton, including: Using the global path skeleton as the centerline of the path corridor, the corridor half-width is calculated based on the target robot's body width and lateral safety margin. The path corridor is generated based on the centerline and the corridor half-width.

10. The semantically constrained path planning and dynamic obstacle avoidance method for quadruped robots according to claim 8, characterized in that, Determine whether the target robot's action was successfully executed. If the determination result is negative, update the semantic reliability, including: Calculate the semantic conflict index based on the execution result of the action and the semantic risk vector; Determine whether the semantic conflict index is greater than the trigger threshold. If the determination result is yes, update the semantic reliability.

Images