An autonomous obstacle avoidance method for a quadrupedal inspection robot for power storage, the robot and the medium

By using graph neural network models and risk field models, the quadruped robot has achieved the prediction of the intention probability of dynamic obstacles and the classification of collision risks in the power storage environment. This solves the problems of accuracy and safety in obstacle avoidance in the power storage environment and ensures the continuity and efficiency of inspection tasks.

CN123113879APending Publication Date: 2026-11-17CHUXIONG POWER SUPPLY BUREAU OF YUNNAN POWER GRID CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611428917.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-15
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Quadruped robots face the challenge of obstacle avoidance in complex mixed environments of dynamic and static obstacles in power storage facilities. Existing obstacle avoidance methods are insufficient to achieve accurate, smooth, and safe autonomous inspection and dynamic obstacle avoidance.

Method used

A graph neural network model is used to predict obstacle trajectories. A set of probabilistic trajectories is generated by combining uniform velocity and uniform angular velocity models. Collision risk is quantified by a risk field model, and a comprehensive obstacle avoidance strategy is generated to achieve probability prediction of the intention of dynamic obstacles and classification of collision risk.

Benefits of technology

It enables quadruped robots to cooperate with dynamic obstacles in real time in power storage environments, ensuring the continuity and efficiency of inspection tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113879A_ABST
    Figure CN123113879A_ABST
Patent Text Reader

Abstract

This application discloses an autonomous obstacle avoidance method, robot, and medium for a quadruped inspection robot in power warehouses, relating to the field of automated warehouse inspection technology. The method establishes a motion model of the quadruped robot in a dynamic obstacle environment within a power warehouse; it uses a uniform velocity model and a uniform angular velocity model to predict the trajectory of the motion model, generating two sets of probabilistic trajectories for future time periods; it then uses a graph neural network model for obstacle avoidance scenarios to predict obstacle trajectories based on interactive awareness, obtaining multimodal trajectories and intent probabilities for future time periods; based on the modality with the highest confidence among the multimodal trajectories, it uses a risk field model to quantify the collision risk value for each future time period, and performs collision risk classification; finally, based on the collision risk level and intent probability, it generates a comprehensive obstacle avoidance strategy for each future time period. This application enables a quadruped inspection robot to perform accurate, smooth, and safe autonomous inspection and dynamic obstacle avoidance in a high-speed, dynamic warehouse environment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of automated warehouse inspection technology, and in particular to an autonomous obstacle avoidance method, robot and medium for a quadruped inspection robot for power warehouses. Background Technology

[0002] With the development of robotics technology, quadruped robots, thanks to their excellent leg mechanisms, possess high mobility and terrain adaptability, and have begun to be applied to inventory and inspection in power warehousing. However, in complex warehousing environments, common unstructured features also bring considerable challenges to quadruped robot inspection operations, mainly in the following aspects: (1) Environmental complexity: The warehouse indoor environment has a high degree of mixed structured and unstructured characteristics. Static obstacles (shelves, pallets, equipment) are regular in shape but densely arranged, and the passage is narrow; dynamic obstacles (personnel, forklifts) have uncertain movement trajectories and are interactive.

[0003] (2) Real-time performance and safety of dynamic obstacle avoidance: The motion control of a quadruped robot is complex and computational resources are limited. When sharing space with moving forklifts or personnel, the obstacle avoidance system must have extremely high real-time performance (millisecond-level response) and predictive capabilities. Simple reactive obstacle avoidance can easily lead to the robot getting stuck in local oscillations or making dangerous movements that violate dynamic constraints.

[0004] (3) Coordination between inspection tasks and obstacle avoidance: The robot needs to avoid obstacles while performing core inspection tasks (such as scanning shelf QR codes and identifying equipment status). This requires that the obstacle avoidance strategy cannot be isolated, but needs to be deeply coupled with global path planning, inspection tasks, etc. Otherwise, it may lead to task interruption or low execution efficiency.

[0005] Therefore, there is an urgent need for a new obstacle avoidance method that can adapt to the special environment of power storage, identify dynamic obstacles, predict short-term movement trajectories, and analyze interactive intentions to achieve accurate, smooth, and safe autonomous inspection and dynamic obstacle avoidance. Summary of the Invention

[0006] The purpose of this application is to provide an autonomous obstacle avoidance method, robot, and medium for a quadruped inspection robot in power storage, which enables the quadruped inspection robot to perform accurate, smooth, and safe autonomous inspection and dynamic obstacle avoidance in a high-speed dynamic storage environment.

[0007] To achieve the above objectives, this application provides the following solution: Firstly, this application provides an autonomous obstacle avoidance method for a quadrupedal inspection robot in power storage facilities, including: Establish a motion model of a quadruped robot in a power warehouse indoor inspection scenario involving dynamic obstacles; The motion model is predicted using a uniform velocity model and a uniform angular velocity model respectively, generating two sets of probabilistic trajectories for future time periods; Based on two sets of probabilistic trajectories, a graph neural network model for obstacle avoidance scenarios of quadruped robots is used to predict obstacle trajectories based on interactive awareness, obtaining the multimodal trajectories of dynamic obstacles in future time periods, the confidence level of each modality, and the probability of intent. In the graph neural network model, the nodes of the graph include the quadruped robot and dynamic and static obstacles, and the edges of the graph include the spatial and semantic relationships between potential interactive nodes. Based on the modal trajectory with the highest confidence among the multimodal trajectories and the quadruped robot's planned trajectory in the future time period, a risk field model is used to obtain the collision risk value at each future time. Based on the collision risk value, a collision risk classification is performed to obtain the collision risk level; Based on the collision risk level and the probability of intent, a comprehensive obstacle avoidance strategy is generated for each future moment.

[0008] Optionally, a motion model of the quadruped robot over dynamic obstacles in an indoor inspection scenario of a power storage facility is established, including: Construct a scenario for quadruped robots to conduct inspections inside a power storage facility; Based on the state of dynamic obstacles in the work scenario, a motion model of the dynamic obstacles is established as follows: ;in, For dynamic obstacles in time The state, ( ) represents the position of the dynamic obstacle. , () represents the components of the dynamic obstacle's velocity in the x and y directions. θ The direction of movement of a dynamic obstacle.

[0009] Optionally, a uniform velocity model and a uniform angular velocity model are used to predict the trajectory of the motion model, respectively, to generate two sets of probabilistic trajectories for future time periods, including: The uniform velocity model is set as follows: In the formula, For dynamic obstacles based on the assumption of linear motion in future time The predicted state, For dynamic obstacles in time state, It is a linear state transition matrix. For noise; The uniform angular velocity model is set as follows: In the formula, For dynamic obstacles based on the turning motion assumption in future time The predicted state, It is a nonlinear state transition function; Based on the uniform velocity model, Kalman filtering is used to predict the state of dynamic obstacles, obtaining the mean and covariance of the predicted trajectories based on the linear motion assumption in the future time period, which constitute a set of probabilistic trajectories for the future time period. ;in, and These are dynamic obstacles in the future time frame based on the assumption of linear motion. and future time The mean of the predicted trajectory, and These are dynamic obstacles in the future time frame based on the assumption of linear motion. and future time covariance; Based on the uniform angular velocity model, Kalman filtering is used to predict the state of dynamic obstacles, obtaining the mean and covariance of predicted trajectories based on the turning motion assumption in the future time period, which constitute another set of probabilistic trajectories for the future time period. ;in, and These are dynamic obstacles in the future time, based on the assumption of turning motion. and future time The mean of the predicted trajectory, and These are dynamic obstacles in the future time, based on the assumption of turning motion. and future time The covariance.

[0010] Optionally, the graph neural network model includes a multi-task output head; for each modality: the multi-task output head is used to simultaneously regress the trajectory coordinates of the modality and output an unnormalized log confidence score, and inputs the log confidence scores of all modalities into the Softmax function for normalization to obtain the intention probability; the multimodalities include continue straight, avoid to the right, and avoid to the left; The node features in the graph neural network model include state, type, and historical trajectory fragments; the types include quadruped robots, dynamic obstacles, and static obstacles.

[0011] Optionally, the historical trajectory segments are obtained by continuously observing dynamic obstacles through sensors and using a multi-target tracking algorithm.

[0012] Optionally, based on the modal trajectory with the highest confidence among the multimodal trajectories and the quadruped robot's planned trajectory for future time periods, a risk field model is used to obtain collision risk values ​​for each future time period, including: The trajectory at each future moment in a modal trajectory is defined by its position, velocity, and orientation. Based on the velocities at each future time step in the modal trajectory with the highest confidence level, the formula is used. and Determine the major and minor axes; where, a For the long axis, a 0 is the base semi-axis. α For the first velocity gain, For velocity modulus, b For the short axis, b 0 is the base short semi-axis, β This is the second velocity gain; Using the dynamic obstacle body center at each future moment in the modal trajectory with the highest confidence as the origin and the motion direction at each future moment as the horizontal axis, and combining the major axis and minor axis, an ellipse in the obstacle body coordinate system is established, and the equation of the ellipse in the obstacle body coordinate system is determined. By using a rotation matrix, the planned trajectory of the quadruped robot at each future moment is transformed from the world coordinate system to the obstacle body coordinate system, thus obtaining the position of the quadruped robot at each future moment in the obstacle body coordinate system; Based on the position of the quadruped robot at each future moment in the obstacle's body coordinate system, the effective distance is calculated using the ellipse equation in the obstacle's body coordinate system. Calculate the relative velocities of the dynamic obstacle and the quadruped robot at each future moment based on the velocities at each future moment in the modal trajectory and the planned velocities of the quadruped robot at each future moment; Based on the effective distance and the relative speed, using the risk function To obtain the collision risk value at each future moment; where, For quadruped robots In the future The collision risk value that is borne For risk amplitude gain, For effective distance, This is the gain coefficient for the relative velocity term. This refers to relative velocity.

[0013] Optionally, it also includes: When the inspection operation scenario includes multiple dynamic obstacles, the collision risk values ​​of all dynamic obstacles at the same future time are superimposed to obtain the total collision risk value at each future time. When the total collision risk value is less than the first threshold value, the collision risk level is determined to be low risk level; When the total collision risk value is less than the second threshold and greater than or equal to the first threshold, the collision risk level is determined to be medium risk; where the first threshold is less than the second threshold. When the total collision risk value is greater than or equal to the second critical value, the collision risk level is determined to be high risk level; the collision risk level increases sequentially from low risk level to medium risk level to high risk level.

[0014] Optionally, when the collision risk level is low, there is no need to adjust the path planning of the quadruped inspection robot. When the collision risk level is medium risk, a comprehensive obstacle avoidance strategy is adopted, which combines the probability of intent, speed control, movement direction adjustment and obstacle avoidance path generation. When the collision risk level is high, adjust the path planning of the quadruped inspection robot.

[0015] Secondly, this application provides a quadrupedal inspection robot for power storage, comprising: the quadrupedal inspection robot being used to perform the autonomous obstacle avoidance method for quadrupedal inspection robots for power storage as described in any one of the above-mentioned methods.

[0016] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the autonomous obstacle avoidance method for a quadrupedal inspection robot for power storage as described above.

[0017] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides an autonomous obstacle avoidance method, robot, and medium for a quadrupedal inspection robot in power storage. The graph neural network model incorporates not only the quadrupedal robot itself and dynamic obstacles, but also static obstacles that significantly influence interaction decisions. The edges encode the spatial and semantic relationships between potential interaction nodes, constructing the entire obstacle avoidance prediction problem on a heterogeneous scene graph. Through the message passing mechanism of the graph neural network, information is iteratively exchanged between the quadrupedal robot, dynamic obstacles, and static obstacles, ensuring that the state encoding of each entity incorporates the contextual information of its neighbors. This achieves a leap from isolated trajectory prediction to scene-contextual awareness and collective behavior understanding. Furthermore, the graph neural network model performs obstacle trajectory prediction based on interaction awareness, identifying the intent probability of dynamic obstacles, enabling the quadrupedal inspection robot to adopt more targeted and collaborative strategies. Simultaneously, this application can generate comprehensive obstacle avoidance strategies for various future moments. This online real-time decision-making technology allows the robot to cooperate seamlessly and safely with dynamic obstacles in a high-speed, dynamic storage environment, effectively ensuring the safety of the quadrupedal inspection robot's movement. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating an autonomous obstacle avoidance method for a quadrupedal inspection robot in power storage, provided in an embodiment of this application; Figure 2 A schematic diagram illustrating the principle of an autonomous obstacle avoidance method for a quadrupedal inspection robot in power storage, provided in an embodiment of this application; Figure 3 This is a schematic diagram of message passing and information aggregation in a graph neural network provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] While existing technologies offer general obstacle avoidance solutions for quadruped robots, most are designed for open outdoor terrain or known static indoor environments. There is a lack of efficient and safe dynamic obstacle avoidance strategies specifically tailored to the power storage scenario, which integrates the handling of dense static obstacles and dynamic obstacles with potential cooperative / conflict intentions. Some industry researchers have proposed Group Interaction Field (GIF) models, which analyze pedestrian trajectories, gaze directions, and group relationships to quantify future positions and attention directions, thus improving the ability of unmanned systems to predict pedestrian intentions to some extent. Others have proposed using Graph Attention Networks (GAT) to dynamically capture interactions between vehicles and introduce uncertainty features to enhance prediction robustness. Still others output multiple possible future obstacle avoidance paths and their probabilities (multimodal probability prediction), using Bayesian Graph Neural Networks (BGNN) to quantify prediction uncertainty and prioritize potential obstacle threats. In addition, some studies have introduced interpretable chain-of-thought (CoT) reasoning mechanisms and visual attention maps to clearly show which environmental factors (such as pedestrian approach or sudden braking of the vehicle in front) led to the current obstacle avoidance decision.

[0023] Because power storage rooms are densely packed environments with a mix of static and dynamic obstacles, existing simple "perception-avoidance" reactive logic has significant limitations in such environments, and can even lead to safety accidents. Quadruped robots can only make reasonable decisions and upgrade from "seeing obstacles" to "understanding the scene and intent" by fully recognizing obstacle motion models, predicting the movement trends of other dynamic objects through short-term extrapolation, and understanding their intentions (whether it's normal passage, yielding, or preparing for work). Without accurate intent understanding, quadruped robots will adopt overly conservative strategies when encountering dynamic obstacles, such as frequent sudden stops or unnecessary detours, which will seriously affect the continuity and efficiency of inspection tasks. Through predictive modeling, robots can plan smooth, efficient, and uninterrupted paths in advance, achieving "unobstructed" obstacle avoidance, thus ensuring the smooth operation of the entire warehouse inspection process.

[0024] In view of this, in an exemplary embodiment, such as Figure 1 As shown, an autonomous obstacle avoidance method for a quadrupedal inspection robot for power storage is provided, which includes the following steps 101 to 106.

[0025] Step 101: Establish a motion model of the quadruped robot in the indoor inspection scenario of the power storage room, which is a dynamic obstacle.

[0026] Step 102: Use the uniform velocity model and the uniform angular velocity model to predict the trajectory of the motion model, respectively, to generate two sets of probabilistic trajectories for future time periods.

[0027] Step 103: Based on the two sets of probability trajectories, use the graph neural network model of the quadruped robot obstacle avoidance scenario to predict the obstacle trajectory based on interactive awareness, and obtain the multimodal trajectory of the dynamic obstacle in the future time period, the confidence of each modality trajectory and the intention probability; the nodes of the graph in the graph neural network model include the quadruped robot and dynamic and static obstacles, and the edges of the graph include the spatial relationship and semantic relationship between potential interactive nodes.

[0028] Step 104: Based on the modal trajectory with the highest confidence among the multimodal trajectories and the quadruped robot's planned trajectory for future time periods, the collision risk value at each future time is obtained using a risk field model.

[0029] Step 105: Based on the collision risk value, determine the collision risk level and obtain the collision risk grade.

[0030] Step 106: Generate a comprehensive obstacle avoidance strategy for each future moment based on the collision risk level and the probability of intent.

[0031] Figure 2The following is a brief overview of the working process of the method described in this application: S1 Establishing a motion model of dynamic obstacles in a power storage facility (corresponding to step 101) → S2 Using a uniform velocity or uniform angular velocity model to perform short-term extrapolation of the obstacle's motion model (corresponding to step 102) → S3 Constructing a graph neural network model to predict obstacle trajectories based on interactive awareness (corresponding to step 103) → S4 Constructing a risk field model to quantify and calculate the collision risk value brought by the obstacle (corresponding to step 104) → S5 Performing collision risk classification and discrimination to support obstacle avoidance decision generation (corresponding to steps 105 and 106). The method of this application follows a progressive logic of "physical prediction → intent reasoning → risk quantification → classification decision," with each step tightly coupled. S1 Establishing the kinematic state vector of the obstacle. (Including position, velocity, and orientation), providing basic input for subsequent predictions; this state vector is directly substituted into the uniform / uniform angular velocity motion model of S2, and the state transition matrix is ​​used to... and nonlinear functions Perform recursion and generate the future using Kalman filtering. The set of probability trajectories for seconds (Including mean and covariance). This set of trajectories. This is then used as the initial features of the nodes in the S3 graph neural network. (Encoding location, speed, type) and edge features The core input (encoding relative position and relative velocity) is used by the network through multi-round message passing (messages). After polymerization After updating the node state, decode and output the multimodal trajectory. and the probability of intent The risk field model of S4 directly uses the predicted trajectory output from step three—utilizing the future position. Calculate the effective distance At the same time, the relative velocity is calculated using the predicted velocity. Substituting both into the risk function Quantify collision risk; the risks of each obstacle are added together to form the total risk. S5 substitutes this total risk into the step discriminant function. The risk level is determined by comparing the results with thresholds N1 and N2, resulting in three levels: low, medium, and high. Simultaneously, the intent probability output from step three (e.g., 72% yield probability) serves as the decision context in this step, guiding the selection of a specific obstacle avoidance strategy (gentle deceleration with a small deviation is adopted if the risk is medium and the yield probability is high; emergency braking is employed if the risk is high). This forms a complete data chain of "state → trajectory → intent → risk → decision," with each step's output serving as a crucial input for the next step, ensuring real-time safe obstacle avoidance.

[0032] In another exemplary embodiment of this application, step 101 above constructs a working scenario for a quadruped robot to inspect an electrical storage room; for dynamic obstacles (taking pedestrians or forklifts as examples), its kinematic model is first established. On a two-dimensional plane (storage floor), the state of a dynamic obstacle in time... It can be represented as: ; The above formula represents the motion model of the dynamic obstacles established based on their state in the work scenario. Wherein, For dynamic obstacles in time The state, ( ) represents the position of the dynamic obstacle. , () represents the components of the dynamic obstacle's velocity in the x and y directions. The direction of movement of a dynamic obstacle.

[0033] In another exemplary embodiment of this application, step 102 above uses a uniform velocity or uniform angular velocity model to perform short-term extrapolation on the motion model of the obstacle established in step one, as described in steps 2.1 to 2.4.

[0034] Step 2.1, set the uniform velocity model as follows: ; ; In the formula, For dynamic obstacles based on the assumption of linear motion in future time The predicted state, It is a linear state transition matrix. To predict the time step, For noise, This represents unmodeled factors such as acceleration.

[0035] Step 2.2, assume the obstacle has an angular velocity ω Motion, its state is: Here, the two velocity components are combined into a single resultant velocity scalar. (i.e., speed), and added angular velocity. As a state variable; here It represents the resultant velocity of the obstacle's movement (i.e., velocity modulus), not... or The mathematical relationship between them is: .Right now, , It is to calculate here. Basic data, It is a scalar value after fusion. The uniform angular velocity model assumes that the object is making circular or turning motion, in which case it uses... (rate) and ω Describing motion using angular velocity is better than using... , More intuitive and with simpler mathematical derivations (e.g., when updating position). Such items).

[0036] The uniform angular velocity model is set as follows: In the formula, For dynamic obstacles based on the turning motion assumption in future time The predicted state, It is a nonlinear state transition function.

[0037] .

[0038] when ω When ≈0, a uniform velocity model can be used as an approximation.

[0039] Step 2.3: Based on the uniform velocity model, Kalman filtering is used to predict the state of dynamic obstacles, obtaining the mean and covariance of the predicted trajectories based on the linear motion assumption in the future time period, and forming a set of probabilistic trajectories for the future time period: ; in, and These are dynamic obstacles in the future time frame based on the assumption of linear motion. and future time The mean of the predicted trajectory, and These are dynamic obstacles in the future time frame based on the assumption of linear motion. and future time The covariance.

[0040] Step 2.4: Based on the uniform angular velocity model, Kalman filtering is used to predict the state of dynamic obstacles, obtaining the mean and covariance of the predicted trajectories based on the turning motion assumption in the future time period, which constitute another set of probabilistic trajectories for the future time period. ; in, and These are dynamic obstacles in the future time, based on the assumption of turning motion. and future time The mean of the predicted trajectory, and These are dynamic obstacles in the future time, based on the assumption of turning motion. and future time The covariance.

[0041] The general expression for the two sets of probability trajectories above is: ; in: N = T / Δ t . To predict the total number of steps, Predict the duration for the future (e.g., 3.0 seconds). The filter update period is 0.1 seconds.

[0042] P It is a set of probability trajectories (i.e., a complete prediction output packet), which is not a single trajectory but contains future trajectories. N A dataset containing the distribution of all possible states at each time step. For the first The mean of the predicted states at each step is the state of the future time step. The most likely position, velocity, and other state values ​​(such as coordinates) are the mean position of the trajectory. It is the first The covariance matrix of the predicted state at each step represents the range of uncertainty of the predicted value at that moment (i.e., the magnitude and direction of the error ellipse), which is the covariance of the trajectory. Note that the mean... It is the center line of the future trajectory, covariance It is the uncertainty zone surrounding the center line, and together the two constitute a complete probability trajectory.

[0043] For the same tracked dynamic obstacle, Kalman filter predictions based on two motion hypotheses are run in parallel to cover both linear and turning motion possibilities. The specific generation methods are shown in Table 1.

[0044] Table 1 Generation Method

[0045] Trajectory Set In practice, it includes two sets of parallel prediction results (CV branch and CT branch), with each time step corresponding to a state estimate and covariance matrix. The subsequent graph neural network will use these two sets of predictions as prior inputs and utilize interaction information to select or fuse more reasonable modalities.

[0046] Step 102 completed the mapping from the "current state" to the "future probability trajectory". This set of probability trajectories... It is used directly as initial input in graph neural network models. Specifically: the predicted position in the set. Covariance Encoded as graph neural network node features A portion of this data (such as the historical trajectory and movement trend of obstacles) is used as edge features. The relative positions and velocities between different obstacles (calculated from the data in the set) are used as edge features. This is used to describe the interaction relationships between nodes. In general, step 102 provides "individual motion priors based on physical laws" for step 103, and step 103 then performs "group trajectory correction based on interaction awareness" on this basis.

[0047] In another exemplary embodiment of this application, in order to predict the trajectory of moving obstacles, a graph neural network model is constructed using the idea of ​​function approximation, and the obstacle avoidance scenario of the quadruped robot is constructed as a graph G=(V,E).

[0048] like Figure 3 As shown, for node V in a graph neural network, each node... The characteristics of a node represent an intelligent agent (robot, dynamic obstacle) or an important scene element (such as shelf end point, aisle entrance). It includes its status (position, speed, type, etc.) and historical trajectory fragments.

[0049] Among them, the "position" and "velocity" in the node features are physically identical to the state vector in step 101. Essentially, they are the optimal motion state estimates of the obstacle or robot at the current moment after filtering and fusion. The graph neural network directly calls this estimate as the initial input of the node attributes. However, the two differ slightly in their data roles. Step 101 provides the original state parameters, while the graph neural network fuses them with neighbor interaction information and then re-encodes them. Therefore, although they have the same numerical source, they are at different levels in the computation graph. As for "type," it is not a kinematic physical quantity, but a discrete semantic label used to distinguish the category attributes of different entities in the graph. It usually includes dynamic obstacles, the quadruped robot itself, and key static structures. The role of this label is to enable the graph neural network to identify "who is who" when passing messages, thereby applying differentiated physical constraints and interaction weights to neighbors of different categories. For example, the robot needs to avoid a forklift but does not need to avoid the fixed shelf endpoint, which is only used as a spatial reference.

[0050] "Historical trajectory fragments" refer to a sequence of actual states stored continuously within several frames prior to the current moment, reflecting the recent motion inertia and path orientation of the entity. In the temporal direction, they represent deterministic observations from the past to the present. This historical data is not generated by the prediction in step 102, but rather obtained through continuous sensor observation and caching by a multi-target tracking algorithm. The system maintains a fixed-length sliding window in memory, storing the latest state and discarding the oldest state with each new frame of data received, thus providing a reliable temporal context for the graph neural network. It is important to emphasize that these historical trajectory fragments and the probabilistic trajectory set output in step two are diametrically opposed in the temporal direction. The former represents established facts from the past to the present, while the latter represents probabilistic predictions from the present to the future. They play different roles in subsequent applications: historical fragments are used to encode the entity's behavioral inertia, while the future predictions in step 102 serve as physical prior inputs, fused with or residually connected to the interactive correction results output by the graph neural network, jointly supporting the generation of the final multimodal trajectory. Therefore, the input of a graph neural network essentially includes three layers of information: the current state, historical inertia, and physical extrapolation prior. The model integrates these with the interaction relationships with its neighbors through message passing, and finally outputs the future trajectory and intent probability after scene semantic correction.

[0051] For edge E in a graph neural network, it is achieved by connecting pairs of nodes that have potential interactions ( , ), edge characteristics It can encode information such as relative position and speed.

[0052] The process of predicting obstacle trajectories uses forward propagation of a graph neural network for inference, as detailed below: 1) Node encoding: A multilayer perceptron is used to map the raw features of each node to a high-dimensional embedding space. .

[0053] Here, in the formula Representative node The original feature vector is precisely the numerical concatenation result of the "original features" of each node described above before any network mapping, containing state information such as the node's position, velocity, and type, as well as fragments of its historical trajectory. Therefore, here... Essentially, it is the raw input data received by the encoder input layer, which, along with the node features mentioned above, They correspond completely in physical terms. It is the specific tensor form of the original features, while It is a conceptual general term for this feature, and both refer to the same original information that has not been transformed.

[0054] and This involves using the aforementioned original feature vectors. The initial node embedding obtained after nonlinear mapping by a multilayer perceptron (MLP), where the superscript... Specifically, this refers to the layer 0 state of a graph neural network before message passing begins, i.e., the initial layer that has not yet exchanged information with any neighboring nodes. The physical meaning of this embedding is precisely the distributed feature representation in a high-dimensional semantic space—it maps low-dimensional physical coordinates, velocity scalars, and discrete type labels to a higher-dimensional, more expressive vector space, enabling the originally scattered heterogeneous raw data to be computed and compared at a unified semantic level. The features belong to the high-dimensional embedding space, rather than the state values ​​of the original physical space; the embedding result provides an abstracted and purified initial expression of the nodes for subsequent multiple rounds of message passing, laying the feature foundation for interaction reasoning between nodes.

[0055] 2) Message passing (multi-round): Perform L rounds of message passing to allow nodes to exchange information.

[0056] Message passing requires message computation for each edge ( , ), calculate from arrive Message: .

[0057] In the formula and These represent the target node (receiver). and the source node (sender) In the graph neural network Embedding of features for the current round after round message passing.

[0058] Specifically, the subscript letters clearly indicate the direction of message passing—the calculation is designed to generate the message from the node. (Sender) Flow to Node The message from the receiver requires both the receiver's current embedding state and the sender's current embedding state to determine what information the sender should transmit to the receiver. (The superscript...) This emphasizes that these two embeddings are not the original initial inputs. It has already passed. The intermediate state after round of iteration updates—this means for In this situation, they have already aggregated their respective... The interaction information within the skip neighborhood includes a certain level of scene context awareness. During this calculation process, and Along with the edge features connecting this pair of nodes Vectors are concatenated along the feature dimension to form a complete input vector, which is then fed into the multilayer perceptron. This allows the network to comprehensively consider three aspects: "who is sending the message," "to whom," and "what is the relationship between them," ultimately outputting a message specifically tailored to the receiving node. Personalized message tensor In short, this concatenation operation makes message generation no longer isolated and unidirectional, but fully combines the dynamic characteristics of the sender and receiver and their relative relationship, providing a precise basis for subsequent aggregation and node updates.

[0059] In addition, message aggregation must be performed during message passing, and the nodes must be processed. Aggregate all from its neighbors The message is as follows: ; AGGREGATE can be a mean, a maximum, or a summation pool.

[0060] 3) Node update: Update the node status with aggregated messages.

[0061] During the forward propagation of a graph neural network, nodes are dynamically updated: or ; here, Represents a node In the The updated node feature embedding after each round of message passing (i.e., the next level). It is derived from the old embedding of the node in the current round. With the neighbor messages just aggregated The result obtained by fusing the common inputs into a GRU or MLP_node is essentially a new state representation of the node after absorbing a round of neighbor interaction information. It contains a broader range of contextual awareness than the previous round (from...). Jump Neighborhood Extension (Skip neighborhood). Through such iterative updates, node features are gradually integrated into the structured semantic information of the surrounding environment, providing a basis for the final decoding output.

[0062] 4) Output: The final embedding of each node after L rounds of iteration. It contains its own context-aware information, which is then input into the decoder as follows: .

[0063] about and Relationship: yes The final special case in the iteration rounds. Represents any intermediate round of message passing in a graph neural network ( ),and This is the preset total number of iteration rounds. When the iteration reaches the last round... At that time, the node embedding obtained after this round of updates is precisely Therefore, both are essentially expressions of "the same state sequence at different iteration stages"— Generally refers to the first The state after the wheel, Specifically refers to the final state after the last round.

[0064] Regarding the relationship between this formula and the previous update formula: the node update formula in the previous text gives the result from the 1st... Wheel state Recursively to the th Wheel state Iteration rules (using GRU or MLP_node to merge and aggregate messages) The system starts from its initial state. Starting from there, repeatedly apply this update formula for a total of Next (i.e.) The node features are iteratively optimized round by round, and the final update is completed exactly when the node features are obtained. Immediately afterwards, the current "output" formula directly outputs this final state. As a decoder The input is mapped to generate a multimodal trajectory. In short, the previous update formula defined the iterative process of "how to move from the current step to the next step," while the current output formula takes the endpoint of the iteration (…). The input is used to complete the final trajectory decoding. The two are sequentially related as "iterative evolution process" and "final state utilization".

[0065] in, It is a node (Dynamic Obstacles) Future The trajectory of each modality, This refers to the number of predicted obstacle modes (e.g., "go straight", "turn left", "turn right"). The network also outputs the confidence score for each obstacle mode. .

[0066] There is a progressive relationship between the trajectories obtained in steps 102 and 103, namely, "physical prior input and scene semantic refinement output". The probabilistic trajectory sets obtained in steps 2.3 and 2.4 are also relevant. The individual physical extrapolation results generated based on the motion model and Kalman filtering only reflect the linear or turning motion assumptions of the obstacle itself and do not contain interactive information; while the multimodal trajectory output in step 103 Based on this, As part of the initial features of the nodes, multiple discrete trajectory candidates are decoded and generated after fusing neighbor interaction information through a graph neural network. Step 102 provides the baseline value and uncertainty covariance, and step 103 performs semantic correction on the physical baseline, outputting a multimodal trajectory with intent confidence to serve subsequent decision-making.

[0067] The process of obtaining it: In the decoder design in step 103, It doesn't just output trajectory coordinates; instead, it's designed as a multi-task output head: for each modality... The decoder simultaneously regresses the trajectory coordinates of this mode. It then outputs an unnormalized log-score confidence level (logit). Subsequently, it processes all... The logit values ​​of each modality are input into the Softmax function for normalization, thus obtaining the result that satisfies... The probability distribution of each This represents the predicted probability that the modality is a true future trajectory. The training of this confidence score relies on a joint loss function—simultaneously optimizing the regression error (e.g., mean squared error) of the trajectory coordinates and the cross-entropy loss of the modality classification—forcing the network to learn to distinguish different behavioral patterns in its internal representation, thereby outputting a confidence score with practical reference value.

[0068] This confidence level plays a key role primarily in step 104 (risk field construction) and steps 105-106 (obstacle avoidance decision generation): Mode selection in step 104: When multiple predicted trajectory modes exist for the same obstacle, the system selects the mode based on the confidence level. The mode with the highest probability is selected as the "most likely scenario" to calculate the future collision risk (as in the example). Since the confidence level is the highest, mode 2 is selected for risk prediction to avoid redundant calculations of the risk field for all modes and ensure real-time performance. At the same time, the confidence level can also be used as a weight to perform probability-weighted fusion of the risk values ​​of each mode to generate a more robust comprehensive risk estimate.

[0069] The decision-making basis in steps 105-106: In the obstacle avoidance decision generation stage, confidence level and intention probability together constitute a comprehensive assessment of prediction reliability. If the highest confidence level is significantly higher than other modalities (e.g., If the confidence level is high (e.g., the probability of each modality is similar), the decision-making system has a high degree of confidence in the prediction and can take targeted actions. Conversely, if the confidence level is dispersed (e.g., the probabilities of each modality are similar), the decision-making system tends to adopt a conservative strategy (e.g., slowing down and observing) to cope with the uncertainty of the prediction. In addition, the confidence level can also be linked to the risk level. For example, under the same risk level, obstacles with higher confidence levels can be prioritized, thereby improving the robustness and safety of obstacle avoidance decisions.

[0070] In another exemplary embodiment of this application, step 104 above models the influence of dynamic obstacles as a risk field, which is a scalar field C. risk (p,t), where p=(x,y) is a spatial point and t is time. This is an asymmetric, time-varying potential field. Step 104 can be replaced by steps 4.1 to 4.7.

[0071] Step 4.1: Define the trajectory of each future moment in the modal trajectory, including position, velocity, and orientation.

[0072] Step 4.2: Based on the velocities at each future time step in the modal trajectory with the highest confidence level, use the formula... and Determine the major and minor axes; where, a For the long axis, a 0 is the base semi-axis. α For the first velocity gain, For velocity modulus, b For the short axis, b 0 is the base short semi-axis, β This represents the second velocity gain. This means that the faster the speed, the farther the risk field extends in the direction of motion.

[0073] Step 4.3: Using the dynamic obstacle body center at each future moment in the modal trajectory with the highest confidence as the origin and the motion direction at each future moment as the horizontal axis, and combining the major axis and minor axis, establish an ellipse in the obstacle body coordinate system, and determine the equation of the ellipse in the obstacle body coordinate system.

[0074] The equation of the ellipse is: ; It is the coordinate of point p in the world coordinate system transformed into the coordinate system of the obstacle itself. Representing any point in space, using coordinates This indicates the location of the risk to be assessed.

[0075] Step 4.4: Using a rotation matrix, the planned trajectory of the quadruped robot at each future moment is transformed from the world coordinate system to the obstacle body coordinate system, thereby obtaining the position of the quadruped robot at each future moment in the obstacle body coordinate system.

[0076] Step 4.5: Based on the position of the quadruped robot at each future moment in the obstacle body coordinate system, calculate the effective distance using the ellipse equation in the obstacle body coordinate system.

[0077] For a point outside the ellipse, the distance is calculated by mapping it to the nearest point on the ellipse, that is: ; For the effective distance, in the current d eff When the value is ≤1, it means the point is inside the ellipse, which poses an extremely high risk.

[0078] Step 4.6: Calculate the relative velocities of the dynamic obstacle and the quadruped robot at each future moment based on the velocities at each future moment in the modal trajectory and the planned velocities of the quadruped robot at each future moment.

[0079] Regardless of spatial point Whether a point is inside or outside the ellipse, the system will simultaneously calculate the effective distance from the point to the ellipse and the relative velocity at the current moment, which together determine the risk value of the point—the elliptic term is responsible for characterizing the "spatial proximity" (positional risk), and the velocity term is responsible for characterizing the "urgency brought about by the approach velocity" (momentum risk).

[0080] Step 4.7: Based on the effective distance and the relative speed, use the risk function. To obtain the collision risk value at each future moment; where, For quadruped robots In the future The collision risk value that is borne For risk amplitude gain, This refers to relative velocity. This is the gain coefficient (weighting parameter) for the relative velocity term, used to adjust the relative velocity. The degree to which risk values ​​are amplified. The larger the value, the more significant the contribution of relative speed to risk, making the robot more sensitive to obstacles approaching at high speed, thereby avoiding side collisions.

[0081] Because dynamic obstacles are in motion, their position and velocity change over time; therefore, the risk field is time-varying. Used to identify the specific moment corresponding to the calculation of the risk value.

[0082] The physical meaning of collision risk value lies in the fact that the risk value depends not only on "how close to the obstacle" but also on "how fast they are relatively close". Even if two obstacles are at the same distance from the robot, an oncoming obstacle traveling at high speed ( The risk value contributed by a large obstacle will be far greater than that of an obstacle following in the same direction and at the same speed. (Small), thus giving the risk field speed sensitivity, ensuring that the robot can give timely warnings and avoidance responses to high-speed approaching vehicles from the side, which is the core mechanism for avoiding side collisions.

[0083] Therefore, it can be seen from the risk function that the risk field is time-varying, because the position and velocity of the obstacle are derived from its predicted trajectory. The velocity is obtained by numerically differentiating the position (i.e., dividing the position difference between adjacent time steps by the time step), or directly from the velocity component output synchronously by the decoder.

[0084] In another exemplary embodiment of this application, the method further includes: when the inspection operation scenario includes multiple dynamic obstacles, superimposing the collision risk values ​​of all dynamic obstacles at the same future time to obtain the total collision risk value at each future time; when the total collision risk value is less than a first threshold value, determining the collision risk level as low risk level; when the total collision risk value is less than a second threshold value but greater than or equal to the first threshold value, determining the collision risk level as medium risk level; when the total collision risk value is greater than or equal to the second threshold value, determining the collision risk level as high risk level; wherein, the first threshold value is less than the second threshold value; the collision risk level of low risk level, medium risk level and high risk level increases sequentially.

[0085] For a warehouse quadruped robot inspection scenario with multiple obstacles, the total risk field is the sum of the risk fields of each obstacle, that is: ; in, For quadruped robots In the future The total collision risk value.

[0086] For example, based on the quantitative calculation results of the collision risk caused by obstacles to the quadrupedal inspection robot, the collision risk is classified and determined. The step function for collision risk classification is designed as follows: ; In the formula, Collision risk level, It is classified as a low-risk level. It is classified as a medium-risk level. It is classified as a high-risk level. This represents the collision risk value. This is the first critical value. This is the second critical value. . , The threshold value (critical value) used to differentiate risk assessment levels is a preset constant, used to classify risks into low, medium, and high levels. When the inspection scenario includes multiple dynamic obstacles, : Total collision risk value It is the comprehensive risk value after the superposition of risks from multiple obstacles.

[0087] If the collision risk quantification result is less than If the collision risk assessment result is low, no adjustment to the existing path planning is required; if the collision risk quantification result is between and If the collision risk is between [a certain value] and [a certain value], the collision risk assessment result is medium. Further comprehensive obstacle avoidance strategies, including speed control, direction adjustment, and obstacle avoidance path generation, can be implemented to reduce the collision risk, while continuous monitoring of the collision risk should be maintained. If the collision risk quantification result is greater than [a certain value], the collision risk assessment result is medium. If the collision risk is high, the path planning needs to be adjusted immediately to reduce the collision risk.

[0088] In another exemplary embodiment of this application, when the collision risk level is low, there is no need to adjust the path planning of the quadruped inspection robot; when the collision risk level is medium, a comprehensive obstacle avoidance strategy combining speed control, movement direction adjustment, and obstacle avoidance path generation is adopted based on the intent probability; when the collision risk level is high, the path planning of the quadruped inspection robot is adjusted.

[0089] Taking the implementation of an autonomous obstacle avoidance method for a quadrupedal inspection robot in a power warehouse based on dynamic obstacle trajectory prediction and interactive intent understanding strategy as an example, by setting the operation scenario of the quadrupedal robot in a power warehouse inspection, the risk of collision between the quadrupedal robot and typical obstacles (forklifts) during the inspection is judged, providing a direct basis for obstacle avoidance decision-making and ensuring the safety and efficiency of the entire inspection process.

[0090] Examples of implementation methods are as follows: Step 1: Establish a motion model of dynamic obstacles in the power storage facility.

[0091] First, we set up the operational scenario for the quadruped robot's inspection of the power warehouse.

[0092] (1) Environmental configuration includes: warehouse aisle width: 2.0 meters; robot current position: (0,0), orientation: 0° (along the positive direction of the aisle); robot current speed: (0.8, 0) m / s; task target point: (10, 0) (10 meters in front of the aisle).

[0093] (2) Dynamic obstacle (forklift) configuration includes: initial position: (6, 0.5) (6 meters in front of the robot, 0.5 meters to the right); initial velocity: (-1.2, 0) m / s (moving towards the robot); forklift length: 2.0 meters, width: 1.0 meter; maximum forklift acceleration: ±0.8 m / s² 2 .

[0094] (3) Sensor parameter configuration includes: update frequency: 10Hz (Δt = 0.1s); prediction duration: T = 3.0s; prediction steps: N=30.

[0095] For a forklift navigating dynamic obstacles, its kinematic model is first established. The forklift's state vector is: .

[0096] Step 2: Model the motion of the obstacle using a uniform velocity or uniform angular velocity model. A uniform velocity model is used to predict the trajectory of obstacles. Its state transition matrix is: .

[0097] The process noise covariance is: Q=diag([0.01,0.01,0.05,0.05,0.001]).

[0098] Therefore, the prediction results for each step size can be predicted: For k = 1 to 30 (corresponding to 0.1s to 3.0s), we have: The prediction results for the first 5 steps are shown in Table 2.

[0099] Table 2 Prediction results of the first 5 steps

[0100] The key prediction points are: 1.0 second later: Forklift position (4.80, 0.50); 2.0 seconds later: Forklift position (3.60, 0.50); 3.0 seconds later: Forklift position (2.40, 0.50).

[0101] Based on this, the collision risk is assessed as follows: The quadruped robot moves forward at 0.8 m / s, and its position after 3 seconds is (2.4, 0); while the forklift's position after 3 seconds is (2.4, 0.5); the relative distance between the two is 0.5 meters < the safe distance (about 1.0 meter), so a close-range conflict is predicted.

[0102] Step 3: Construct a graph neural network model to predict obstacle trajectories based on interactive awareness.

[0103] To predict the trajectory of moving obstacles, a graph neural network model is constructed using the idea of ​​function approximation, and the obstacle avoidance scenario of the quadruped robot is constructed as a graph G=(V,E).

[0104] (1) Scene graph construction.

[0105] 1) Node characteristics are as follows: Forklift node v1: [6.0, 0.5, -1.2, 0.0, 1, 0, 0] (position x, y, speed v) x ,v y Type=1 (Vehicle), Historical Speed ​​v' x ,v' y ).

[0106] Robot node v2: [0.0, 0.0, 0.8, 0.0, 2, 0.8, 0.0] (type=2(robot)).

[0107] Shelf node v3: [8.0, 1.5, 0, 0, 3, 0, 0] (Type=3 (static structure), identifies the right boundary of the aisle).

[0108] 2) Edge characteristics are as follows: e1→2: Relative position [-6.0, -0.5], relative velocity [2.0, 0.0], distance is 6.02; e2→1: Relative position [6.0, 0.5], relative velocity [-2.0, 0.0], distance is 6.02; e1→3: Relative position [2.0, 1.0], relative velocity [-1.2, 0.0], distance is 2.24.

[0109] (2) GNN forward propagation calculation 1) Coding layer: ; .

[0110] 2) Message passing (round 1): Message computation v2→v1 is as follows: ; The message aggregation is as follows: ; The node updates are as follows: .

[0111] 3) After three rounds of message passing, the final node is embedded: .

[0112] (3) Trajectory prediction output The decoder outputs three modes: Mode 1: Continue straight (probability π1=0.25), maintain the current speed and direction trajectory, and the position will be (2.40, 0.50) after 3 seconds.

[0113] Mode 2: Rightward avoidance (probability π2=0.65); Trajectory: Starts turning right after 1.0 seconds, with a shift in the y-direction; Turning start: t=1.0s, position (4.80, 0.50); ​​Turning process: uniform angular velocity ω=0.3rad / s; Position after 3 seconds: (2.40, 0.80), orientation: 3.04rad.

[0114] Mode 3: Avoid to the left (probability π3 = 0.10); Trajectory: Emergency left turn; Position after 3 seconds: (2.40, 0.20), Orientation: 3.24 rad (4) Intent recognition output: yielding intent: 72% probability; cutting in: 18% probability; ignoring intent: 10% probability.

[0115] Step 4: Construct a risk field model and quantify the collision risk value caused by obstacles.

[0116] The impact of dynamic obstacles is modeled as a risk field. The parameters for this risk field are set as follows: Basic half-axis: a 0 = 2.0 meters b 0 = 1.0 meter; velocity gain: α =0.5, β =0.3; Risk amplitude: k =2.0; Relative velocity gain: γ =0.8.

[0117] (1) Calculation of the risk field at the current moment The ellipse parameters in the obstacle's body coordinate system are calculated as follows: Forklift speed modulus: =1.2m / s; ; .

[0118] The risk calculation for the quadruped robot's location is as follows: A transformation from the world coordinate system to the forklift coordinate system is achieved using a rotation matrix (oriented towards π rad): ; Therefore, the quadruped robot's position (0,0) in the forklift coordinate system is: ; The effective distance is calculated by mapping it to the nearest point on the ellipse as follows: .

[0119] The relative velocities were calculated as follows: robot (0.8, 0), forklift (-1.2, 0); ; .

[0120] The risk value at the current moment can be further calculated: .

[0121] (2) Future risk field prediction Calculate future collision risk based on Mode 2 (most likely scenario): At t=2.0s: the predicted position of the forklift is (3.60, 0.65); the predicted position of the quadruped robot is (1.60, 0.0). Thus, the relative position of the quadruped robot and the forklift is (-2.0, -0.65), which is transformed into the forklift coordinate system as (2.0, 0.65).

[0122] Obtain the relative velocity: (Assuming the speed remains constant).

[0123] Based on this, the predicted risk values ​​for future moments are as follows: ; .

[0124] Step 5: Assess and classify collision risks Let the critical values ​​for distinguishing risk assessment levels be N1=2 and N2=5, then we have: ; in, The evaluation result is given; N1 and N2 are the critical parameters for distinguishing risk assessment and grading.

[0125] Due to the calculated collision risk quantification value Therefore, the collision risk classification and evaluation result is medium.

[0126] Since the predicted risk cost is 2.27 after 2 seconds, the quadrupedal inspection robot adopts a comprehensive obstacle avoidance strategy that includes speed control, movement direction adjustment, and path generation. For example: ① Reduce forward speed: from 0.8m / s to 0.4m / s-0.6m / s.

[0127] ② Consider lateral movement: The quadruped robot shifts to the right.

[0128] ③ Generate avoidance trajectory: Plan a trajectory that deviates 0.2-0.3 meters to the right during the period of 1.5 meters to 2.0 seconds.

[0129] Based on the 72% probability of yielding intent under the intent recognition output condition in step three, the obstacle avoidance strategy is formulated as follows: choose active avoidance rather than emergency stop, gentle deceleration + slight right descent rather than aggressive avoidance, and expect the forklift to cooperate to complete space sharing.

[0130] To examine the effectiveness of the obstacle avoidance strategy, it was verified through field testing. Observations were made in real-world scenarios: t=0.0-1.5s: The quadruped robot gently decelerates to 0.5m / s and begins to veer to the right; t=1.5s: The forklift detects the robot's avoidance action and begins to turn right to avoid it; t=2.0-3.0s: The two parties pass each other in the center of the passage, with a minimum distance of 0.8 meters; After t=3.0s: the quadruped robot resumes its original path and speed; Therefore, the accuracy of the obstacle avoidance method proposed in this application is assessed as follows: (1) Final position error: 0.15 meters (predicted (2.40, 0.80) vs. actual (2.40, 0.75)); (2) Accurate intent recognition: Correctly identify the "yielding intent"; (3) Effective risk warning: Conflict risks are identified 2 seconds in advance.

[0131] The proposed autonomous obstacle avoidance method for quadrupedal inspection robots in power warehouses is based on a dynamic obstacle trajectory prediction and interaction intent understanding strategy. Through physical model-based prediction, a simple uniform or uniformly accelerated motion model is extrapolated for dynamic obstacles within the warehouse to generate a short-term trajectory set. Further, based on the prediction requirements based on interaction awareness, a lightweight graph neural network is used to model the spatial relationships between the quadruped robot, dynamic obstacles, and key static obstacles (such as shelf exits and charging piles). This predicts the most likely trajectory distribution of dynamic obstacles and provides the probability of their "interaction intent" (e.g., yielding, cutting in, ignoring). On this basis, an asymmetric, velocity-adaptive elliptical risk field is constructed for each dynamic obstacle. The risk value increases exponentially with decreasing distance and increasing relative speed. By superimposing the risk fields of each obstacle, a total risk field is obtained, enabling collision risk classification and discrimination. This supports obstacle avoidance decision generation, ultimately promoting the practical application and intelligentization of power warehouse inspection robots.

[0132] In terms of functional enhancement: This application proposes a multimodal probabilistic output framework for joint reasoning of "physical trajectory-interaction intent," achieving a cognitive upgrade from predicting the behavior of warehouse obstacles to recognizing motivations. This framework designs a multi-task learning head at the decoding end of the GNN, simultaneously regressing trajectory coordinates and classifying intent labels. The model's loss function simultaneously optimizes the accuracy of trajectory prediction and the accuracy of intent classification, forcing the network to learn features in its internal representation that can distinguish different decision motivations. This achieves a cognitive upgrade from behavior prediction to motivation recognition. The quadruped robot can not only predict the movement of forklifts but also understand whether their behavior is "friendly yielding" or "emergency avoidance." This provides extremely valuable high-level semantic information for subsequent decision-making layers, enabling the quadruped inspection robot to adopt more targeted and collaborative strategies.

[0133] In terms of performance optimization, this application constructs a scene graph neural network model that integrates spatial topology and semantic relationships, achieving a structured deep understanding of the quadruped robot inspection interaction scenario under complex warehousing conditions. The entire obstacle avoidance prediction problem is constructed on a heterogeneous scene graph. The nodes of this graph not only include the quadruped robot itself and dynamic obstacles, but also incorporate static scene elements that significantly influence interaction decisions. The edges of the graph encode spatial relationships (such as relative position and velocity) and semantic relationships (such as "approaching the exit of the passage") between entities. Through the message passing mechanism of the graph neural network, information is iteratively exchanged in multiple rounds among nodes such as the quadruped robot, personnel, forklifts, and shelves. This allows the state encoding of each entity to incorporate the contextual information of its neighbors, achieving a leap from isolated trajectory prediction to scene-context-aware group behavior understanding, providing support for generating predictions that conform to scene common sense.

[0134] In terms of advantages: This application provides a lightweight, end-to-end trainable architecture that seamlessly embeds interactive prediction into the real-time planning loop of a quadrupedal inspection robot in a warehouse, effectively ensuring safety. In the designed lightweight and end-to-end trainable graph network architecture, the real-time operation capability of the model on the aforementioned computational units is ensured through careful design of the feature dimensions of nodes and edges, control of the number of message passing layers, and the use of efficient aggregation functions. The entire model, from warehouse scene graph construction, feature encoding, message passing to final trajectory and intent decoding, is integrated into an end-to-end neural network. This network takes real-time scene snapshots generated by sensors as input and directly outputs prediction results with uncertainty metrics that can be directly used for motion planning. This innovation makes high-level interactive intent understanding no longer an offline analysis tool, but a core component of online real-time decision-making. It can update predictions at a frequency of 10Hz or even higher, directly embedding the "agent's" thinking model into the control loop of the quadrupedal inspection robot, realizing a closed loop of perception-prediction-decision-action. This enables the robot to cooperate smoothly and safely with other moving entities (including obstacles) in a real-time, dynamic warehouse environment, effectively ensuring the safety of the quadruped inspection robot's movement.

[0135] Based on the same inventive concept, this application also provides a quadrupedal inspection robot for power storage warehouses, used to implement the autonomous obstacle avoidance method of the quadrupedal inspection robot for power storage warehouses described above. The solution provided by this quadrupedal inspection robot is similar to the solution described in the above method. Therefore, the specific limitations of the one or more embodiments of the quadrupedal inspection robot for power storage warehouses can be found in the limitations of the autonomous obstacle avoidance method for quadrupedal inspection robots for power storage warehouses described above, and will not be repeated here. The quadrupedal inspection robot for power storage warehouses here uses a quadrupedal robot with an existing structure. The internal program of the existing quadrupedal robot is improved, and the internal program executes the autonomous obstacle avoidance method of the quadrupedal inspection robot for power storage warehouses described above.

[0136] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described in the embodiment of the autonomous obstacle avoidance method for a quadrupedal inspection robot for power storage.

[0137] 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.

[0138] 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 method for autonomous obstacle avoidance of a quadrupedal inspection robot for power storage, characterized in that, include: Establish a motion model of a quadruped robot in a power warehouse indoor inspection scenario involving dynamic obstacles; The motion model is predicted using a uniform velocity model and a uniform angular velocity model respectively, generating two sets of probabilistic trajectories for future time periods; Based on two sets of probabilistic trajectories, a graph neural network model for obstacle avoidance scenarios of quadruped robots is used to predict obstacle trajectories based on interactive awareness, obtaining the multimodal trajectories of dynamic obstacles in future time periods, the confidence level of each modality, and the probability of intent. In the graph neural network model, the nodes of the graph include the quadruped robot and dynamic and static obstacles, and the edges of the graph include the spatial and semantic relationships between potential interactive nodes. Based on the modal trajectory with the highest confidence among the multimodal trajectories and the quadruped robot's planned trajectory in the future time period, a risk field model is used to obtain the collision risk value at each future time. Based on the collision risk value, a collision risk classification is performed to obtain the collision risk level; Based on the collision risk level and the probability of intent, a comprehensive obstacle avoidance strategy is generated for each future moment.

2. The autonomous obstacle avoidance method for a quadrupedal inspection robot in power storage facilities according to claim 1, characterized in that, Establish a motion model for a quadruped robot in a power storage indoor inspection scenario, including: Construct a scenario for quadruped robots to conduct inspections inside a power storage facility; Based on the state of dynamic obstacles in the work scenario, a motion model of the dynamic obstacles is established as follows: ;in, For dynamic obstacles in time The state, ( ) represents the position of the dynamic obstacle. , () represents the components of the dynamic obstacle's velocity in the x and y directions. θ The direction of movement of a dynamic obstacle.

3. The autonomous obstacle avoidance method for a quadrupedal inspection robot in power storage facilities according to claim 1, characterized in that, The motion model is predicted using both a uniform velocity model and a uniform angular velocity model, generating two sets of probabilistic trajectories for future time periods, including: The uniform velocity model is set as follows: In the formula, For dynamic obstacles based on the assumption of linear motion in future time The predicted state, For dynamic obstacles in time state, It is a linear state transition matrix. For noise; The uniform angular velocity model is set as follows: In the formula, For dynamic obstacles based on the turning motion assumption in future time The predicted state, It is a nonlinear state transition function; Based on the uniform velocity model, Kalman filtering is used to predict the state of dynamic obstacles, obtaining the mean and covariance of the predicted trajectories based on the linear motion assumption in the future time period, which constitute a set of probabilistic trajectories for the future time period. ;in, and These are dynamic obstacles in the future time frame based on the assumption of linear motion. and future time The mean of the predicted trajectory, and These are dynamic obstacles in the future time frame based on the assumption of linear motion. and future time covariance; Based on the uniform angular velocity model, Kalman filtering is used to predict the state of dynamic obstacles, obtaining the mean and covariance of predicted trajectories based on the turning motion assumption in the future time period, which constitute another set of probabilistic trajectories for the future time period. ;in, and These are dynamic obstacles in the future time frame based on the assumption of turning motion. and future time The mean of the predicted trajectory, and These are dynamic obstacles in the future time frame based on the assumption of turning motion. and future time The covariance.

4. The autonomous obstacle avoidance method for a quadrupedal inspection robot in power storage according to claim 1, characterized in that, The graph neural network model includes a multi-task output head; for each modality: the multi-task output head is used to simultaneously regress the trajectory coordinates of the modality and output an unnormalized log confidence score. The log confidence scores of all modalities are input into the Softmax function for normalization to obtain the intention probability; the multimodalities include continue straight, avoid to the right, and avoid to the left; The node features in the graph neural network model include state, type, and historical trajectory fragments; the types include quadruped robots, dynamic obstacles, and static obstacles.

5. The autonomous obstacle avoidance method for a quadrupedal inspection robot in power storage according to claim 4, characterized in that, The historical trajectory segments are obtained by continuously observing dynamic obstacles through sensors and using a multi-target tracking algorithm.

6. The autonomous obstacle avoidance method for a quadrupedal inspection robot in power storage facilities according to claim 1, characterized in that, Based on the modal trajectory with the highest confidence among the multimodal trajectories and the quadruped robot's planned trajectory for future time periods, a risk field model is used to obtain the collision risk values ​​at each future time point, including: The trajectory at each future moment in a modal trajectory is defined by its position, velocity, and orientation. Based on the velocities at each future time step in the modal trajectory with the highest confidence level, the formula is used. and Determine the major and minor axes; where, a For the long axis, a 0 is the base semi-axis. α For the first velocity gain, For velocity modulus, b For the short axis, b 0 is the base short semi-axis, β This is the second velocity gain; Using the dynamic obstacle body center at each future moment in the modal trajectory with the highest confidence as the origin and the motion direction at each future moment as the horizontal axis, and combining the major axis and minor axis, an ellipse in the obstacle body coordinate system is established, and the equation of the ellipse in the obstacle body coordinate system is determined. By using a rotation matrix, the planned trajectory of the quadruped robot at each future moment is transformed from the world coordinate system to the obstacle body coordinate system, thus obtaining the position of the quadruped robot at each future moment in the obstacle body coordinate system; Based on the position of the quadruped robot at each future moment in the obstacle's body coordinate system, the effective distance is calculated using the ellipse equation in the obstacle's body coordinate system. Calculate the relative velocities of the dynamic obstacle and the quadruped robot at each future moment based on the velocities at each future moment in the modal trajectory and the planned velocities of the quadruped robot at each future moment; Based on the effective distance and the relative speed, using the risk function To obtain the collision risk value at each future moment; where, For quadruped robots In the future The collision risk value that is borne For risk amplitude gain, For effective distance, This is the gain coefficient for the relative velocity term. This refers to relative velocity.

7. The autonomous obstacle avoidance method for a quadrupedal inspection robot in power storage according to claim 1, characterized in that, Also includes: When the inspection operation scenario includes multiple dynamic obstacles, the collision risk values ​​of all dynamic obstacles at the same future time are superimposed to obtain the total collision risk value at each future time. When the total collision risk value is less than the first threshold value, the collision risk level is determined to be low risk level; When the total collision risk value is less than the second threshold and greater than or equal to the first threshold, the collision risk level is determined to be medium risk; where the first threshold is less than the second threshold. When the total collision risk value is greater than or equal to the second critical value, the collision risk level is determined to be high risk level; the collision risk level increases sequentially from low risk level to medium risk level to high risk level.

8. The autonomous obstacle avoidance method for a quadrupedal inspection robot in power storage according to claim 7, characterized in that, When the collision risk level is low, there is no need to adjust the path planning of the quadruped inspection robot. When the collision risk level is medium risk, a comprehensive obstacle avoidance strategy is adopted, which combines the probability of intent, speed control, movement direction adjustment and obstacle avoidance path generation. When the collision risk level is high, adjust the path planning of the quadruped inspection robot.

9. A quadruped inspection robot for power storage, characterized in that, The quadruped inspection robot is used to perform the autonomous obstacle avoidance method for a quadruped inspection robot in power storage as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the autonomous obstacle avoidance method for the quadrupedal inspection robot for power storage as described in any one of claims 1-8.