Robot Dynamic Obstacle Avoidance Path Planning Method and System

Abstract

This invention discloses a method and system for dynamic obstacle avoidance path planning for robots. The method includes: collecting data such as images and UWB tag coordinates from multiple mobile robots via an edge computing terminal, and using this data to construct a dynamic environment spatiotemporal database; preprocessing the collected images; identifying and classifying static and dynamic obstacles; outputting their feature information and motion vectors; obtaining the reliable weights of each robot; predicting the future position and trajectory of dynamic obstacles; forming predicted risk areas; using static landmarks to correct positioning errors; dynamically adjusting obstacle avoidance strategies and path planning; generating collaborative scheduling instructions to guide local obstacle avoidance or global task reallocation; and recording detailed execution logs. By implementing the method of this invention, robot swarms can achieve high-precision, low-latency, continuous, and safe autonomous obstacle avoidance and path planning in complex dynamic environments.

Description

Technical Field

[0001] This invention relates to robot control methods, and more specifically to a robot dynamic obstacle avoidance path planning method and system. Background Technology

[0002] With the large-scale deployment of mobile robots in warehousing and logistics, public services, and intelligent inspection, the ability of robot swarms to navigate safely and collaborate in complex dynamic environments has become a key bottleneck restricting their application depth. Existing technologies primarily rely on local perception devices such as LiDAR and depth cameras mounted on individual robots for obstacle detection and path planning. This results in significant perception blind spots and spatiotemporal limitations, especially when dealing with suddenly appearing dynamic obstacles (such as pedestrians or other mobile robots), lacking effective motion intent prediction mechanisms. This leads to delayed obstacle avoidance decisions, frequent path interruptions, and even collision risks. Furthermore, the isolated perception information between multiple robot systems makes cross-terminal data complementarity and collaborative decision-making difficult. In scenarios with dense crowds or severe obstacle obstruction, the uncertainty of single-robot perception increases significantly, severely impacting overall scheduling efficiency and safety.

[0003] In recent years, the development of multi-sensor fusion and edge computing technologies has provided new ideas for collaborative perception in robots. Some studies have attempted to construct a global environment map by centrally processing perception data uploaded by multiple robots in the cloud to achieve collaborative obstacle avoidance. However, this centralized architecture faces problems such as high communication latency, large data redundancy, and poor system robustness. Meanwhile, although dynamic obstacle trajectory prediction methods based on multi-target tracking have been applied in the field of autonomous driving, when directly transferred to robotic scenarios, there is still a lack of online evaluation and reliable measurement mechanisms for perception data quality, which cannot effectively address false detections and missed detections caused by changes in lighting, sensor noise, or algorithm defects. More importantly, existing technologies have failed to fully explore the anchoring role of static environmental features such as building corners and fixed facilities in robot localization and state verification, and have not established a complete closed-loop process from data verification, trajectory prediction, state determination to behavior control, resulting in a lack of continuity and foresight in the robot's obstacle avoidance response in dynamic environments.

[0004] Therefore, it is necessary to design a new method that can fully utilize the advantages of multi-robot collaborative perception, integrate cross-terminal data spectrum verification and spatiotemporal trajectory prediction, and make intelligent decisions based on credibility weighting, in order to solve core technical problems such as inaccurate prediction of dynamic obstacle intent, large differences in the reliability of multi-robot data, positioning drift and state misjudgment, and realize high-precision, low-latency, continuous and safe autonomous obstacle avoidance and path planning of robot swarms in complex dynamic environments. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for dynamic obstacle avoidance path planning for robots.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a robot dynamic obstacle avoidance path planning method, characterized in that it includes: The system acquires images, UWB tag coordinates, timestamps, and body motion state data collected by edge computing terminals mounted on multiple mobile robots, and constructs a dynamic environment spatiotemporal database based on the collected data. The image is preprocessed to identify and classify static obstacles, dynamic obstacles, and temporary occlusion targets. The obstacle feature information, confidence score, and motion vector are output to obtain the multi-robot collaborative perception results. Furthermore, the cross-terminal data in the dynamic environment spatiotemporal database is subjected to interference removal and validity verification using spectral similarity calculation and correlation coefficient verification methods to obtain the corresponding confidence weights for each robot. Based on the multi-robot collaborative perception results, dynamic obstacles close to the robot's planned path are selected and their future positions are predicted. The dynamic obstacle's trajectory is constructed by associating cross-robot, continuous multi-frame data, and calibrated in a unified world coordinate system to generate a predicted risk area, thus forming an obstacle spatiotemporal trajectory prediction result. The obstacle spatiotemporal trajectory prediction result is associated with the timestamps in the spatiotemporal database and the UWB tag coordinates. In the preprocessed image, a feature point detection algorithm is used to identify and extract fixed static landmarks. The relative position vector of the target robot with respect to the static landmarks is calculated based on the UWB label coordinates. The robot position fingerprint is constructed to correct the positioning error and verify the stationary state. Based on the obstacle spatiotemporal trajectory prediction results, the robot position fingerprint, the multi-robot collaborative perception results, and the corresponding trusted weights of each robot, the obstacle avoidance strategy on the planned path is dynamically determined and the path is replanned to generate collaborative scheduling instructions. Based on the collaborative scheduling instructions, and combined with the robot's current state and environmental characteristics retrieved from the spatiotemporal database, local obstacle avoidance behavior control or global task reallocation is performed, and a structured execution log is generated.

[0007] This invention also provides a robot dynamic obstacle avoidance path planning system, comprising: The acquisition unit is used to acquire images, UWB tag coordinates, timestamps and body motion state data collected by edge computing terminals mounted on multiple mobile robots, and to build a dynamic environment spatiotemporal database based on the acquired data; The processing unit is used to preprocess the image, identify and classify static obstacles, dynamic obstacles and temporary occlusion targets, output obstacle feature information, confidence score and motion vector, obtain multi-robot collaborative perception results, and use spectral similarity calculation and correlation coefficient verification methods to perform interference removal and validity verification on cross-terminal data in the dynamic environment spatiotemporal database to obtain the corresponding confidence weight of each robot. The prediction unit is used to filter dynamic obstacles close to the robot's planned path and predict their future positions based on the multi-robot collaborative perception results. It constructs the dynamic obstacle's motion trajectory by associating cross-robot, continuous multi-frame data, and calibrates it in a unified world coordinate system to generate a predicted risk area, thus forming an obstacle spatiotemporal trajectory prediction result. The obstacle spatiotemporal trajectory prediction result is associated with the timestamps in the spatiotemporal database and the UWB tag coordinates. The fingerprint construction unit is used to identify and extract fixed static landmarks in the preprocessed image using a feature point detection algorithm, calculate the relative position vector of the target robot relative to the static landmarks based on the UWB label coordinates, and construct the robot position fingerprint to correct positioning errors and verify the stationary state. The instruction generation unit is used to dynamically determine the obstacle avoidance strategy and replan the path based on the obstacle spatiotemporal trajectory prediction result, the robot position fingerprint, the multi-robot collaborative perception result, and the corresponding trust weight of each robot, and generate collaborative scheduling instructions. The execution unit is used to perform local obstacle avoidance behavior control or global task reallocation based on the cooperative scheduling instructions and the robot's current state and environmental characteristics retrieved from the spatiotemporal database, and to generate a structured execution log.

[0008] The advantages of this invention compared to existing technologies are as follows: This invention constructs a dynamic spatiotemporal database by collecting environmental data from edge computing terminals mounted on multiple mobile robots, and verifies the validity and reliability weights of cross-terminal data using spectral similarity and correlation coefficients, thereby solving the problem of inconsistent data reliability among multiple robots. This method preprocesses images to identify static and dynamic obstacles, and combines continuous multi-frame data from across robots to predict the future position and trajectory of dynamic obstacles, forming accurate spatiotemporal trajectory prediction results for obstacles, thus improving the accuracy of dynamic obstacle intention prediction. Simultaneously, static landmarks are used to correct positioning errors, improving positioning accuracy and solving the positioning drift problem. Based on the above information and the reliability weights of each robot, the system can make intelligent decisions, dynamically adjust obstacle avoidance strategies and path planning, generate collaborative scheduling instructions, guide local obstacle avoidance or global task reallocation, and ensure that the robot cluster achieves high-precision, low-latency, continuous, and safe autonomous obstacle avoidance and path planning in complex dynamic environments. This comprehensive method fully utilizes the advantages of multi-robot collaborative perception, achieving reliable data fusion and intelligent decision-making.

[0009] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart illustrating the robot dynamic obstacle avoidance path planning method provided in an embodiment of the present invention; Figure 2 This is a schematic block diagram of a robot dynamic obstacle avoidance path planning system provided in an embodiment of the present invention; Figure 3 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present 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 the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0014] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0015] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0016] Please see Figure 1 , Figure 1 This is a flowchart illustrating the robot dynamic obstacle avoidance path planning method provided in this embodiment of the invention. By constructing a multi-robot dynamic environment spatiotemporal database, and using spectral verification and feature fusion to remove data interference and assign reliable weights, dynamic obstacle trajectories and predicted risk areas are constructed based on the filtered effective data. Simultaneously, robot position fingerprints are constructed by extracting static landmarks to correct localization. Then, hierarchical obstacle avoidance decisions and path replanning are performed by integrating the trajectory prediction, position fingerprints, and reliable weights from the preceding steps. Finally, control is executed and logs are generated. A tight data flow loop is formed between each step, achieving high-precision collaborative obstacle avoidance and continuous path planning for robot swarms in complex dynamic environments.

[0017] Figure 1 This is a flowchart illustrating the robot dynamic obstacle avoidance path planning method provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps S110 to S160.

[0018] S110. Acquire images, UWB tag coordinates, timestamps, and body motion state data collected by edge computing terminals mounted on multiple mobile robots, and construct a dynamic environment spatiotemporal database based on the collected data.

[0019] This step is the foundational data collection and fusion stage of the entire robot dynamic obstacle avoidance path planning method. It aims to collect multi-source heterogeneous data in real time through multiple distributed mobile robots, and after spatiotemporal alignment processing, to build a unified, complete, and reliable dynamic environment spatiotemporal database, providing data support for subsequent obstacle recognition, trajectory prediction, conflict assessment, and collaborative scheduling.

[0020] In one embodiment, step S110 described above may include steps S111 to S112.

[0021] S111. Real-time acquisition of RGB images, depth images, UWB tag coordinates, timestamps, and the robot's linear velocity, angular velocity, heading angle data, current battery level, load weight, and motion state data through the edge computing devices of each robot to obtain detection data; among which, the motion state data includes maximum speed, maximum acceleration, and braking distance.

[0022] In this embodiment, this step utilizes edge computing terminals deployed on each mobile robot to collect the robot's own state and environmental perception data in real time at a high frequency, forming a raw detection dataset. The specific data collected is as follows: RGB image: Visible light image acquired by a high-definition color camera on the robot, containing color information from the red, green, and blue channels. It is used to identify the visual features of obstacles such as color, texture, and shape, and is the basis for subsequent semantic recognition and fine-grained feature extraction.

[0023] Depth images are pixel-level distance information images acquired by depth cameras (such as structured light, ToF, or binocular vision systems). Each pixel value represents the spatial distance from that point to the camera. They are used to accurately measure the three-dimensional position, size, and relative distance of obstacles to the robot, making up for the lack of spatial information in RGB images.

[0024] UWB tag coordinates: These are the robot's three-dimensional spatial coordinates (X, Y, Z) or two-dimensional planar coordinates (X, Y) in a local positioning coordinate system, obtained through a UWB (Ultra-Wideband) positioning module. They are used to determine the robot's absolute position in indoor or enclosed environments. Specifically, ranging communication is achieved between UWB base stations (anchors) deployed in the environment and UWB tags mounted on the robot. Positioning is calculated based on TOF (Time of Flight) or TDOA (Time Difference of Arrival) algorithms, achieving centimeter-level (10-30 cm) accuracy. This is a key benchmark for achieving spatiotemporal alignment of multi-robot data and constructing a unified world coordinate system.

[0025] Timestamp: A precise time stamp (usually accurate to milliseconds) that records the moment each data is collected. It is used to synchronize data from different robots and different sensors on the same event and to build a time-consistent trajectory data chain.

[0026] Linear velocity: The instantaneous velocity of the robot body in the forward direction, measured in meters per second. It reflects the current speed of the robot's movement and is a core parameter for subsequent calculation of collision time and prediction of motion trajectory.

[0027] Angular velocity: The instantaneous angular rate at which the robot body rotates around its vertical axis, measured in radians per second. It reflects the robot's current turning speed and is used to predict its heading trend.

[0028] Heading angle data: The angle value of the robot's current direction of motion relative to the geographic North Pole or a custom coordinate system, usually expressed in 0-360 degrees or 0-2π radians. It is used to determine the robot's orientation and is necessary information for constructing motion vectors and trajectory predictions.

[0029] Current battery level: The percentage of the robot's remaining battery power or remaining runtime is a key indicator for assessing the robot's ability to work continuously and the feasibility of tasks. Robots with low battery levels will be assigned low-risk tasks or prioritized for relocation to charging areas.

[0030] Load weight: The weight of the object currently carried by the robot, which affects its maximum speed, acceleration and braking performance, and is an important basis for dynamically adjusting motion parameters and obstacle avoidance strategies.

[0031] Body motion state data includes maximum speed (the highest speed the robot can reach under current load and power conditions), maximum acceleration (the upper limit of the robot's ability to accelerate from a standstill to its maximum speed), and braking distance (the distance required for the robot to brake in an emergency at its current speed). These parameters directly determine the robot's motion constraints and safety boundaries, and are used in path planning and conflict assessment to calculate the minimum safe distance and the feasibility of emergency obstacle avoidance.

[0032] In addition, the local positioning coordinate system is established with the UWB main base station deployed at the edge of the work area as the origin and the long side of the work area as the positive X-axis, forming a right-handed coordinate system; the UWB tag coordinates of all robots are transformed to this unified coordinate system to eliminate the cumulative error of the local coordinate system of a single robot.

[0033] S112. The UWB base station coordinate matching algorithm is used to perform spatiotemporal alignment and association of the detection data of different robots passing through the same location, forming a dynamic environment spatiotemporal database under a unified coordinate system based on the UWB base station network. The dynamic environment spatiotemporal database includes obstacle positions, static landmark coordinates, robot pose, power load status, body motion status data and data confidence scores.

[0034] This step, based on the raw multi-robot detection data collected by S111, uses algorithms to eliminate spatiotemporal inconsistencies and construct a structured global database: UWB base station coordinate matching algorithm: This algorithm identifies whether robots have passed through the same geographical location in physical space by comparing the UWB tag coordinates reported by different robots (usually within an error range of 1-3 meters). When multiple robots are detected arriving at the same location at similar times, a data association mechanism is triggered, marking their perception data as observations of the same environmental area, providing a prerequisite for subsequent spectral similarity calculation and consistency verification.

[0035] Spatiotemporal alignment and association: Due to potential microsecond-level deviations in the internal clocks of each robot, and the susceptibility of UWB ranging to multipath effects and non-line-of-sight (NLOS) propagation resulting in decimeter-level errors, timestamp calibration and spatial coordinate transformation are necessary to unify the data of all robots onto the same global time axis and world coordinate system. Time alignment employs high-precision clock synchronization of the UWB base station network itself (wired or wireless synchronization protocols with nanosecond-level accuracy) or supplemented by the NTP protocol to ensure that each robot's timestamps are based on the same clock reference. Spatial association, through unified coordinate mapping based on the UWB base station coordinate system, directly associates the local coordinates measured by each robot via UWB tags to the global coordinate system. NLOS detection and elimination algorithms are used to eliminate abnormal ranging values, ensuring that the same obstacle "seen" by different robots has a unique and consistent identifier in the database.

[0036] Unified coordinate system: Establish a global reference coordinate system (such as a planar coordinate system with the center of the site as the origin and the north direction as the Y-axis). The positions of all robots, obstacles, and landmarks are all transformed to represent them in this coordinate system, eliminating data heterogeneity caused by different robot orientations and starting positions. This is the spatial basis for subsequent trajectory prediction, conflict calculation, and collaborative scheduling.

[0037] Dynamic Environment Spatiotemporal Database: A real-time updated, structured data center, organized using a spatiotemporal index (timestamp + spatial coordinates) as its core organizational method, storing and managing all detection data and processing results uploaded by robots. This database not only records snapshots of the current moment but also retains historical trajectory data, supporting time window queries and trend analysis.

[0038] Obstacle location: In a unified coordinate system, the three-dimensional coordinates (X, Y, Z) of each detected static or dynamic obstacle. For dynamic obstacles, a velocity vector is also included. This is the direct input for path avoidance and collision detection.

[0039] Static landmark coordinates: Long-term fixed feature point coordinates in the environment, such as corners of walls, columns, and fixed equipment, are used as reference anchor points for robot localization. This is used to construct position fingerprints and correct jump errors in UWB localization caused by non-line-of-sight propagation (NLOS) or multipath effects.

[0040] Robot pose: The precise position (X, Y) and orientation (heading angle) of each robot in a unified coordinate system constitute complete pose information, which is used for multi-robot relative position calculation and cooperative strategy formulation.

[0041] Battery load status: The current battery percentage and load weight of each robot are extracted from S111 and used to assess task execution capability and risk tolerance.

[0042] Data confidence score: Based on the quality of the raw data collected by S111 (such as image sharpness, UWB ranging signal-to-noise ratio (SNR) and ranging accuracy) and the verification results of S112 (such as spectral consistency and correlation coefficient), a confidence value between 0 and 1 is assigned to each data entry. Data with high confidence scores have higher weight in subsequent decisions.

[0043] Specifically, the working principle of the UWB base station coordinate matching algorithm is as follows: Each robot's UWB tag calculates its own coordinates using a TOF (Time of Flight) ranging algorithm. Assume there are N UWB base stations deployed in the environment, with coordinates as follows: The distances measured by the UWB tags on robot j to each base station are... The label coordinates can be determined using trilateration or least squares methods. The UWB tag coordinates of robot j in a unified coordinate system were calculated. The coordinate accuracy can reach 10-30 centimeters.

[0044] For any two robots j and k, the UWB tag coordinates collected at time t are respectively and Calculate the Euclidean distance between them: .

[0045] Set spatial distance threshold =2.0 meters (based on a typical UWB positioning accuracy of 30cm, 3-6 times higher, ensuring correct correlation within a 95% confidence interval). When At that time, it was initially determined that the two robots were in the same physical location.

[0046] Simultaneously set a time window threshold. Seconds, through NTP protocol or the nanosecond-level clock synchronization of the UWB base station network itself, ensure that the data timestamps t of robots j and k are synchronized. j and t k satisfy .

[0047] When the spatial proximity condition is met simultaneously ( (meters) and time synchronization conditions ( At a certain time (seconds), the data association mechanism is triggered, marking the perception data (images, point clouds, etc.) of robots j and k at that moment as collaborative observations of the same environmental area, assigning them the same spatial association identifier, which is used for subsequent spectral similarity calculation and cross-terminal verification.

[0048] S120. The image is preprocessed to identify and classify static obstacles, dynamic obstacles, and temporary occlusion targets. Obstacle feature information, confidence scores, and motion vectors are output to obtain multi-robot collaborative perception results. Furthermore, the cross-terminal data in the dynamic environment spatiotemporal database is subjected to interference removal and validity verification using spectral similarity calculation and correlation coefficient verification methods to obtain the corresponding credible weights for each robot.

[0049] In this embodiment, the multi-robot collaborative perception result refers to the structured environment description dataset generated after four layers of processing from steps S121 to S125. Specifically, it includes three core types of information: First, obstacle feature information, including the category label (static / dynamic / temporary occlusion), geometric attributes (boundary box size, center point coordinates, contour point set), visual attributes (color histogram, texture descriptor), and persistent identifiers (first detection time, number of continuous tracking frames) for each obstacle; second, confidence score, a 0-1 range value assigned to each obstacle target, comprehensively reflecting multi-dimensional quality indicators such as detection stability, feature clarity, and cross-robot consistency, directly determining the target's influence weight in subsequent trajectory prediction and decision-making; and third, motion vectors, for dynamic obstacles and temporary occlusions, including their velocity vectors (vx, vy) in a unified world coordinate system, acceleration estimates, and a sequence of predicted trajectory points within the next 3-5 seconds. This result is the sole input source for obstacle trajectory prediction in step S130, and its data quality directly determines the performance ceiling of the entire obstacle avoidance system.

[0050] The confidence weight for each robot refers to the 0-1 interval weight coefficient dynamically assigned to each robot node in step S123 through statistical correlation analysis. This weight quantifies the reliability level of the robot's perceived data at the current moment. This weight is not statically fixed but dynamically adjusted based on the deviation between the real-time calculated environmental illumination-image grayscale correlation coefficient and the group mean. When a robot's sensor experiences image quality degradation due to occlusion, dirt, or hardware failure, its correlation coefficient will significantly deviate from the cluster statistical distribution, and the system automatically reduces its weight to below 0.3, significantly weakening its contribution to trajectory fitting in step S130. Conversely, robots with stable performance maintain a weight above 0.8, and their detection data is considered high-confidence observations. This weight matrix serves as a key parameter for the collaborative scheduling decision in step S150, directly affecting the fairness and safety of multi-robot task allocation and behavior control.

[0051] This step is the core processing stage for in-depth processing and quality control of the raw sensing data, based on the dynamic environment spatiotemporal database constructed in step S110. Its execution logic follows a four-layer progressive architecture: "first purify the data, then verify the quality, then fuse features, and finally generate the result." This aims to solve common problems in multi-robot collaborative sensing, such as data heterogeneity, noise interference, and confidence differences. Through spectral domain consistency verification and statistical correlation analysis, the system can automatically identify and suppress the influence of unreliable robot nodes. Simultaneously, it utilizes deep learning technology to extract multi-scale features from the purified data, ultimately outputting cross-validated high-confidence collaborative sensing results and a quantifiable robot weight system.

[0052] Specifically, for robots j and k that are determined to be associated with the same location through step S122, the temporal grayscale signal of the target obstacle region in their respective continuously acquired image sequences is extracted: Data length (number of sampling points): N=10 frames, corresponding to a time window T=1.0 second (assuming the camera frame rate is 10fps). Sampling frequency: f s =10Hz, which satisfies the Nyquist sampling theorem (the frequency of obstacle motion is usually <5Hz). Signal extraction method: For each frame of the image, extract the average gray value g(n) within the bounding box of the target obstacle, n=0,1,...,9 Grayscale signal sequence of robot j: The grayscale signal sequence of robot k: ; Performing an N-point FFT on the grayscale signal sequence yields its frequency domain representation: ; Spectral resolution: ; Take the amplitude spectrum (ignore the phase, as the phase is inconsistent due to differences in the viewpoints of different robots): ; Effective frequency range: This corresponds to the frequency of the obstacle's movement.

[0053] To eliminate the difference in absolute signal energy, the amplitude spectrum is normalized: Constructing spectral feature vectors: ; The cosine similarity method is used to calculate the similarity of the spectral feature vectors of the two robots: ; Alternatively, the Pearson correlation coefficient can be used: ;in This is the mean.

[0054] Set a similarity threshold S th=0.7 (Empirical value, which can be determined through ROC curve optimization): like If the spectrum is determined to be consistent, the data from both robots is valid and will be retained for subsequent fusion; if : If it is determined to be a interference signal, outlier data with low similarity will be removed.

[0055] In one embodiment, step S120 described above may include steps S121 to S125.

[0056] S121. Perform denoising and enhancement standardization preprocessing on the image to eliminate the effects of sensor noise and uneven illumination, and obtain a standardized image sequence that conforms to a unified resolution and color space.

[0057] In this embodiment, this step performs an initial purification process to address the inherent quality defects in the raw image data acquired in step S110. Due to differences in camera models, installation angles, and exposure parameters among the various robots, and the dynamic changes in lighting conditions in the actual operating environment (such as indoor-outdoor transition zones and shadow occlusion), the raw images often suffer from noise interference, uneven brightness, and inconsistent resolution. The specific processing flow includes: firstly, smoothing the image using Gaussian filtering or bilateral filtering algorithms to eliminate sensor thermal noise and quantization noise; then, using adaptive histogram equalization (CLAHE) technology to enhance image contrast, making details in both dark and bright areas clearly visible; subsequently, through perspective transformation and image cropping, uniformly correcting the perspective of all robots to a standard top view or front view to eliminate geometric distortion caused by installation angles; finally, performing scaling operations to normalize the image resolution to standard sizes such as 1280×720 or 640×480, and converting the color space to the sRGB standard to ensure strict consistency of the input data received by the subsequent deep learning model. This processing enables images from different robots to achieve homogeneity at both the geometric and optical levels, laying a solid foundation for cross-robot feature comparison.

[0058] S122. For the detection data of different mobile robots on the same obstacle in the same time window in the dynamic environment spatiotemporal database, calculate the similarity metric value of the spectrum response curve, identify outlier data with similarity below the preset threshold as interference signals and remove them, and retain only the valid detection data with spectrum consistency to obtain valid data.

[0059] In this embodiment, this step implements cross-robot collaborative verification at the data level to address false alarms caused by sensor malfunctions, environmental glare, or algorithmic misdetection. When multiple robots generate detection responses to the same spatial region (the spatial distance threshold calculated based on the coordinates of each robot's UWB tag is set to 2 meters, and the accuracy of spatial association is ensured by utilizing UWB centimeter-level positioning accuracy) at similar times (the time window is typically set to ±0.5 seconds), the system extracts the spectral response curve of each robot to the obstacle—that is, the frequency characteristics of the pixel value changes of the obstacle target over time. Specifically, the gray-level mean sequence of the obstacle region in 10 consecutive frames of robot images is subjected to Fast Fourier Transform (FFT) to obtain its frequency domain energy distribution curve. Theoretically, if each robot observes the same real obstacle, its spectral curves should be highly similar (the main energy is concentrated in the low-frequency band, reflecting the slow movement or stationary characteristics of the obstacle); if a robot generates a false target due to image blurring or misidentification, its spectrum will show high-frequency noise or abnormal peaks. The system calculates the cosine similarity or Pearson correlation coefficient between the spectrum curves of each robot, identifies outlier data with a similarity of less than 0.7 as interference signals and removes them from the spatiotemporal database, retaining only the valid detection data that passes the consistency check, thereby significantly reducing the false alarm rate.

[0060] S123. For the valid data, extract the ambient illumination value and the average gray value of the image and calculate the Pearson correlation coefficient between them. For a single robot, the Pearson correlation coefficient measures the internal perception consistency. For a multi-robot cluster, compare the distribution of the Pearson correlation coefficients of each robot. When the coefficient of a certain robot deviates from the group mean by more than the set standard deviation range, the perception credibility is determined to be abnormal and the data weight is reduced to obtain the credibility weight corresponding to each robot.

[0061] In this embodiment, this step assesses the health status of each robot's perception system from the perspective of illumination consistency. Ambient illuminance values ​​are acquired through the robot's onboard illumination sensor, while the average grayscale value of the image is calculated from the global pixels of the image after S121 standardization. Physically, these two should exhibit a strong positive correlation: high illuminance results in a bright image, and low illuminance results in a dark image. For a single robot, the Pearson correlation coefficient between its current illuminance and grayscale values ​​is calculated. If the coefficient is below 0.6, it indicates an anomaly in its camera exposure control or sensor, resulting in poor internal consistency. For a multi-robot cluster, the coefficient distribution of each robot is further compared: the mean μ and standard deviation σ of all robot coefficients are calculated. When a robot's coefficient deviates from the group mean by more than 2σ, it is identified as an abnormal node in perception reliability, and its reliability weight is reduced from the default value of 1.0 to the range of 0.3-0.5, weakening its voice in subsequent data fusion. This weight is dynamically adjusted and recalculated every 5 seconds to ensure the system can adapt to gradual degradation issues such as sensor aging and lens contamination.

[0062] Specifically, the internal consistency threshold for a single robot is r_th = 0.6 (Pearson correlation coefficient); the threshold for group deviation is k = 2 (standard deviation multiple, i.e., the 2σ principle); the weight adjustment range is: normal weight: w_default = 1.0; abnormal weight: w_abnormal = 0.3~0.5 (linear interpolation, the larger the deviation, the lower the weight); and the dynamic adjustment cycle is T_update = 5 seconds.

[0063] S124. The pre-trained large model is called to extract semantic verification features from the standardized image sequence. At the same time, a dual-branch deep learning network is used to detect fine-grained element features that conform to the pixel size range. The two types of features are mapped to a unified high-dimensional feature space and then weighted and fused to generate a fusion feature vector that comprehensively represents the data quality.

[0064] In this embodiment, this step implements deep feature fusion to improve the robustness of target representation from both semantic understanding and detail perception perspectives. A pre-trained large model (such as CLIP, DINOv2, or other visual foundational models) encodes the S121-normalized image, extracting 512-dimensional or 768-dimensional high-level semantic features. These features include obstacle category semantics (e.g., "pedestrian," "forklift," "temporarily stacked boxes"), scene context (e.g., "narrow passage," "intersection"), and potential risk warnings (e.g., "slippery area"), forming semantic verification features. Simultaneously, the detection branch in the dual-branch deep learning network uses a YOLOv8 or DETR architecture to detect small targets in the image with pixel sizes ranging from 32×32 to 256×256, extracting fine-grained element features, including obstacle edge gradients, corner positions, texture patterns, and other low-level geometric details. The feature completion branch then uses contextual information to fill in missing parts for targets with occlusion areas less than 30%, avoiding missed detections. Semantic verification features and fine-grained element features are projected onto a unified 1024-dimensional feature space through a fully connected layer and then fused using a weighted average of 0.6:0.4 to obtain a fused feature vector. This vector contains not only semantic information about "what it is" but also detailed information about "what it looks like," providing a richer representation for subsequent classification and confidence scoring.

[0065] Specifically, the semantic verification feature weight is α=0.6; the fine-grained element feature weight is β=0.4; satisfying α+β=1.0; Feature vector dimensions: Semantic features: 512 or 768 dimensions; Fine-grained features: 256 or 512 dimensions; Unified dimension after fusion: D_fusion = 1024 dimensions.

[0066] S125. Based on the effective data, fused feature vectors, and the confidence weights of each robot, identify and classify static obstacles, dynamic obstacles, and temporary occlusion targets, calculate motion vectors, and generate multi-robot collaborative perception results containing obstacle feature information, confidence scores, and motion vectors.

[0067] In this embodiment, this step is the final output of S120, generating a structured perception output by integrating all previous processing results. Based on the effective data after interference removal in S122 and the fused feature vector generated in S124, a multi-classifier (such as a support vector machine or a lightweight neural network) is used to classify the target in three levels: if the target's position changes by less than 0.1 meters in 20 consecutive frames, it is determined to be a static obstacle (such as a wall or pillar); if the target's position changes continuously and its speed is greater than 0.2 meters per second, it is determined to be a dynamic obstacle (such as a pedestrian or other robot); if the target suddenly appears and has no fixed shape (such as temporarily stacked goods) and does not disappear after 10 frames, it is determined to be a temporary occlusion. For dynamic and temporary targets, Kalman filtering or optical flow is used to calculate their motion vectors (velocity direction and magnitude). The confidence score is a weighted average of three factors: the quality score of the fused feature vector in S124 (0-1), the robot confidence weight assigned in S123 (0-1), and the proportion of frames continuously tracked by the target (0-1). The final confidence score is obtained by weighting these three factors. The final output of the multi-robot collaborative perception result is a list of structures, each element of which contains: target ID, category label, bounding box coordinates, confidence score, motion vector (if any), and a list of source robot IDs. This result is used for trajectory prediction in step S130.

[0068] In this embodiment, the multi-robot collaborative perception result refers to the structured environment description dataset generated after interference removal, feature fusion, and cross-validation. This dataset contains the category attributes, geometric features, motion vectors, and confidence scores of each obstacle, serving as the sole input source for subsequent trajectory prediction and conflict assessment.

[0069] The credibility weight for each robot refers to a reliability coefficient in the range of 0-1, which is dynamically calculated based on the deviation of each robot's illumination-grayscale correlation coefficient from the cluster's statistical distribution. This coefficient is used to quantify the contribution of its perceived data to collaborative decision-making. The higher the value, the more credible the robot's current state is.

[0070] Specifically, static obstacle detection: duration: N_static = 20 frames (2 seconds @ 10fps); position change threshold: Δp_th = 0.1 meters; Dynamic obstacle detection: velocity threshold: v_th = 0.2 m / s; Temporary occlusion detection: Frame count of occurrence: N_temp = 10 frames (1 second @ 10fps); No fixed shape determination: The aspect ratio change rate of the bounding box is >30%.

[0071] S130. Based on the multi-robot collaborative perception results, dynamic obstacles close to the robot's planned path are selected and their future positions are predicted. The dynamic obstacle's trajectory is constructed by associating cross-robot, continuous multi-frame data, and calibrated in a unified world coordinate system to generate a predicted risk area, thus forming an obstacle spatiotemporal trajectory prediction result. The obstacle spatiotemporal trajectory prediction result is associated with the timestamps in the spatiotemporal database and the UWB tag coordinates.

[0072] This step is a crucial bridge connecting environmental perception and obstacle avoidance decision-making. It receives the purified collaborative perception results and robot confidence weights from step S120, and uses kinematic modeling and statistical methods to transform discrete obstacle detection points into continuous, quantifiable spatiotemporal trajectory predictions. The system first filters out dynamic obstacles that pose a potential threat to the robot's path at a global level. Then, it improves the robustness of trajectory estimation through cross-robot observation data fusion, ultimately generating a predicted risk area containing the probability distribution of future positions, providing quantifiable decision input for subsequent conflict assessment.

[0073] In this embodiment, the obstacle spatiotemporal trajectory prediction result refers to a structured spatiotemporal probability model, which includes the predicted trajectory point sequence of each dynamic obstacle within a future time window of 3-5 seconds, the confidence ellipse of each point (reflecting positional uncertainty), and the kinematic parameters of the trajectory (velocity, acceleration, and direction of motion), serving as the direct calculation basis for conflict risk quantification in step S150.

[0074] In one embodiment, step S130 described above may include steps S131 to S135.

[0075] S131. Based on the multi-robot collaborative perception results, identify and select target dynamic obstacles close to the robot's planned path, and associate the detection data of the same obstacle under the perspective of multiple robots to obtain the target obstacle after cross-view association.

[0076] In this embodiment, this step implements threat screening and cross-robot data registration. The system loads the current planned paths of each robot (a sequence of polyline segments provided by the global path planning module), and calculates the shortest Euclidean distance from the current position to the path for each dynamic obstacle in the multi-robot collaborative perception results output in step S125. If this distance is less than the danger distance threshold (usually set as the robot braking distance plus a 1-meter safety margin), the obstacle is determined to be a target dynamic obstacle and included in the subsequent trajectory prediction scope. This mechanism ensures that computing resources are concentrated on targets that truly pose a threat. For the same obstacle observed simultaneously by multiple robots (judged based on the obstacle spatial distance calculated based on the UWB label coordinates of each robot being less than 1 meter and the category labels being consistent), the system initiates cross-view data association: by comparing the obstacle bounding box overlap rate (IoU) and feature similarity provided by different robots, a cross-robot ID mapping table is established to aggregate the scattered observation data into a unified obstacle instance. For example, robot A detects "pedestrian P" at coordinates (10.0, 5.0), while robot B simultaneously detects "pedestrian P" at coordinates (10.2, 5.1) (the above coordinates are two-dimensional planar positions in a unified world coordinate system based on the UWB base station network, in meters). The system confirms that the two are the same target through coordinate system alignment and calibration parameter compensation, merges their observation records, and forms a target obstacle data structure with cross-view association containing complementary information from multiple perspectives, which significantly improves the observation redundancy and anti-occlusion capability of subsequent trajectory estimation.

[0077] Danger distance threshold calculation formula: D_danger = d_brake + d_margin; d_brake: braking distance at the robot's current speed (from S111 body motion state data); d_margin = 1.0 meter (fixed safety margin); cross-view correlation spatial threshold: D_assoc = 1.0 meter (high confidence correlation based on UWB accuracy).

[0078] S132. Analyze the target obstacles in the continuous time-standardized image sequence, including: applying a multi-target tracking algorithm to establish a motion trajectory for each obstacle, and transforming the trajectory to a unified world coordinate system to obtain spatiotemporally standardized trajectory data.

[0079] In this embodiment, this step constructs a historical motion profile of the obstacle. The system backtracks all detection records (approximately 20 consecutive frames) of the target obstacle within the past 2 seconds from the spatiotemporal database in step S110. A multi-target tracking algorithm (such as DeepSORT or ByteTrack) is applied, based on Kalman filter prediction and Hungarian algorithm matching, to establish a temporally continuous motion trajectory for each target. The trajectory consists of a series of state vectors, each containing a timestamp, bounding box position, and velocity estimate. Since the original detection data originates from the local camera coordinate system of each robot, the trajectory points in the local coordinate system need to be transformed to a unified world coordinate system (such as a metric coordinate system with the southwest corner of the site as the origin) using the camera extrinsic parameter matrix and UWB label coordinates calibrated in step S110, eliminating viewpoint differences. The spatiotemporally standardized trajectory data obtained after the transformation is a sequence of world coordinate points with precise time labels, such as [(t1, x1, y1), (t2, x2, y2), ...], with uniform temporal intervals and unified spatial coordinates, providing standardized input for subsequent motion pattern analysis.

[0080] Specifically, the backtracking time window is T_track = 2.0 seconds; Backtracking frame count: N_track = 20 frames (fixed at 10fps); The trajectory point sequence format is: [(t_i, x_i, y_i, v_xi, v_yi)], i=1,2,...,20.

[0081] S133. Analyze the spatiotemporally standardized trajectory data of multiple obstacles in the same scene, calculate the overall motion vector, fit the obstacle motion baseline and expected velocity, and obtain the kinematic prediction model.

[0082] In this embodiment, this step extracts the group movement patterns and individual movement patterns from individual trajectories. The system aggregates the spatiotemporally standardized trajectory data of all target obstacles in the current scene and uses Principal Component Analysis (PCA) to calculate the overall principal direction of motion (i.e., the overall motion vector) of the scene. For example, in a corridor scene, most pedestrians move along the long side of the corridor, and this direction vector is the principal direction of motion. For each target obstacle, its historical trajectory points are orthogonally projected onto the straight line defined by the overall motion vector to obtain a one-dimensional projected coordinate sequence. Based on this sequence, the least squares method is used to fit the obstacle's motion baseline, which represents the centerline of the obstacle's most likely motion path. The displacement difference between adjacent projection points is further calculated, and the time average is obtained to obtain the expected velocity. The kinematic prediction model constructed in this way contains two core parameters: the motion baseline direction vector and the expected velocity scalar. This model assumes that the obstacle maintains uniform linear motion in the short term, which is the basic assumption for extrapolating future positions.

[0083] Specifically, PCA principal component selection: retain principal components with a variance contribution rate > 85%; Least square fitting: RANSAC algorithm was used, with an interior point threshold ε = 0.05 meters; S134. Based on the kinematic prediction model, taking the current obstacle position as the prediction starting point, extrapolate the average movement distance along the motion baseline to determine the world coordinates of the prediction point at the future time.

[0084] In this embodiment, this step performs short-time trajectory extrapolation. The system takes the world coordinates of the target obstacle at the latest time (t0) as the prediction starting point, and extrapolates at a uniform speed along the motion baseline direction according to the kinematic prediction model obtained in step S133. To predict the position at a future time Δt (e.g., Δt = 3 seconds), the average movement distance d = desired speed × Δt is calculated, and the prediction starting point is translated by a distance d along the baseline direction to obtain the world coordinates (x, y, y) of the predicted point at the future time. p y p To address motion uncertainties, the system not only predicts single locations but also generates a sequence of points at multiple future times (e.g., consecutive predicted points at t0+1s, t0+2s, and t0+3s), forming a predicted trajectory. This trajectory is a deterministic prediction based on the current motion state and does not yet include uncertainty quantification; its accuracy is highly dependent on the applicability of the kinematic model.

[0085] Specifically, the prediction time window is: Δt∈{1s,2s,3s,4s,5s} (multi-time prediction); Extrapolation formula: p(t0+Δt)=p(t0)+v̄·Δt·d̂; v̄: desired velocity (calculated in step S133); d̂: unit direction vector of the motion baseline.

[0086] S135. Based on the world coordinates of the predicted point, create a predicted risk region in a unified world coordinate system with the predicted point as the center. The vertical axis is a number of times the position standard deviation and the horizontal axis is a number of times the obstacle size. Transform the predicted risk region to the local coordinate system of each robot to obtain the spatiotemporal trajectory prediction result of the obstacle.

[0087] In this embodiment, this step quantifies the prediction uncertainty and generates an operable decision region. Based on the positional fluctuation statistics of the target obstacle's historical trajectory points, the system calculates its positional standard deviation in the vertical direction of movement (reflecting the degree of lateral sway). In a unified world coordinate system, a prediction risk region is created centered on each prediction point generated in step S134: the longitudinal (along the direction of movement) length of this region is set to 2-3 times the positional standard deviation to cover possible acceleration and deceleration fluctuations; the lateral (perpendicular to the direction of movement) width is set to 1.5 times the maximum size of the obstacle (e.g., if a pedestrian's width is 0.5 meters, then the region width is 0.75 meters) to cover lateral offset risks. This elliptical or rectangular region is essentially a confidence interval for the obstacle's future position, and its area directly reflects the prediction uncertainty. To support distributed control, the system maps the risk region in this world coordinate system back to the local camera coordinate system of each robot through inverse coordinate transformation, generating a personalized risk map. The final output of the obstacle spatiotemporal trajectory prediction results includes: the predicted trajectory point sequence, the vertex coordinates of the predicted risk area polygon corresponding to each point, and the trajectory confidence (based on the goodness of fit of historical data), providing a spatiotemporal probability model that can be directly used for collision detection in step S150.

[0088] Specifically, the longitudinal (direction of motion) length is: L_long = k_long·σ_pos, where k_long = 2.5 (2-3 times the median value). Horizontal (vertical) width: L_lat = k_lat·w_obj, where k_lat = 1.5; σ_pos: Standard deviation of the position of historical trajectory points (vertical direction of movement); w_obj: Maximum size of the obstacle (diagonal of the bounding box or the size of the preset category).

[0089] S140. In the preprocessed image, a feature point detection algorithm is used to identify and extract fixed static landmarks. The relative position vector of the target robot with respect to the static landmarks is calculated based on the UWB label coordinates. The robot position fingerprint is constructed to correct the positioning error and verify the stationary state.

[0090] This step is crucial for improving the robot's self-localization accuracy and motion state self-verification. It aims to address the issues of UWB positioning in complex industrial environments, such as susceptibility to multipath effects, local jumps and cumulative drift caused by non-linear motion (NLOS), and ensuring positioning continuity in areas with UWB base station coverage blind spots or severe electromagnetic interference. By identifying long-term stable geometric feature points in the environment as natural landmarks, the system establishes a dual anchoring mechanism of UWB-vision fusion. Its core logic is that even if UWB coordinates experience instantaneous jumps due to non-line-of-sight propagation or electromagnetic interference, or enter coverage blind spots, the robot's relative position relative to fixed landmarks, obtained through visual recognition, remains accurate and repeatable. By calculating the relative vector with landmarks in real time and fusing it with UWB coordinates for verification, the system not only achieves centimeter-level positioning accuracy but also identifies whether the robot is truly stationary through temporal fingerprint comparison, avoiding misjudging "UWB positioning jitter" or "slipping in place" as valid displacement. This is crucial for subsequent task allocation and obstacle avoidance decisions.

[0091] In this embodiment, the robot position fingerprint refers to a feature descriptor composed of relative position vectors of the target robot relative to 2-3 nearest static landmarks, calculated jointly by the target robot based on UWB coordinates and visual observation. It is represented in multi-dimensional vector form (e.g., [Δx1, Δy1, Δx2, Δy2, Δx3, Δy3]) and is used to uniquely characterize the robot's relative position in the environmental topology. When the UWB signal is stable, this fingerprint and UWB coordinates together construct a robust position estimate; when the UWB signal is occluded or experiences abnormal jumps, this fingerprint serves as an independent positioning reference, calculating the robot's position through geometric constraints and marking UWB data anomalies. This fingerprint is used as a positioning correction input in step S150 and as the core basis for determining stationary status in step S160.

[0092] In one embodiment, step S140 described above may include steps S141 to S143.

[0093] S141. Convert the image into a distortion-free top view using the camera calibration matrix.

[0094] In this embodiment, this step eliminates camera lens distortion and unifies the viewing angle, laying the foundation for accurately measuring the geometric relationship between the robot and the landmark. Because the wide-angle camera on the robot exhibits radial distortion (fisheye effect) and tangential distortion, directly using the original image to calculate the position vector would introduce systematic errors. The system calls the camera intrinsic parameter matrix (including focal length and principal point coordinates) and distortion coefficients (k1, k2, p1, p2) pre-calibrated offline using the Zhang Zhengyou calibration method to perform inverse distortion mapping on the image standardized in step S121, correcting curved lines to straight lines. Furthermore, combining the extrinsic parameter matrix formed by the camera mounting height and pitch angle, a perspective transformation is performed, projecting the tilted image onto a virtual top-view plane (simulating the perspective of a drone shooting vertically downwards). After this transformation, the coordinates of each pixel in the image have a linear correspondence with its actual position in the world coordinate system, simplifying the conversion from pixel distance to real distance to a single scale factor, greatly improving the accuracy of subsequent relative position calculations. The top view output resolution is 800×600, and the pixel resolution corresponds to an actual physical size of 1cm / pixel, providing geometrically accurate input for step S142.

[0095] S142. Use a feature point detection algorithm to identify and extract fixed static landmarks in the top view, including wall corners, column corners, and fixed equipment corners.

[0096] In this embodiment, this step automatically extracts natural landmarks with long-term stability from the corrected top view. The system uses feature point detection algorithms (such as SIFT, ORB, or SuperPoint) to detect corner features with scale invariance and rotation invariance on the top view. Since the top view has eliminated distortion, the corners of structural objects such as corners, columns, and fixed equipment appear as stable L-shaped or T-shaped intersections in the image. The algorithm filters the detected corners: first, it removes candidate points located within the bounding boxes of dynamic obstacles; second, it retains points with fixed material features such as concrete and metal through texture analysis; finally, it uses the spatiotemporal database from step S110 to verify historical persistence, retaining only points with a repeat detection rate of over 90% in the past 24 hours as static landmarks. Each landmark is assigned a unique ID and its precise coordinates in a unified world coordinate system are recorded (e.g., corner point A is located at (15.23m, 8.67m)). This landmark database is effective for a long time in indoor scenes with slow environmental changes, providing a reliable reference benchmark for step S143.

[0097] Specifically, the landmark search distance threshold is: D_landmark = 3.0 meters; Historical persistence validation: P_persist > 90% (repeat detection rate within 24 hours); Top view resolution: 800×600 pixels; Pixel-physical mapping: 1 pixel = 1 centimeter.

[0098] S143. Calculate the relative position vector of the target robot with respect to a static landmark at a distance that meets the requirements based on the UWB tag coordinates, and use it as the robot's position fingerprint. If the difference in position fingerprints at different times is within the set tolerance range and the time interval exceeds the preset minimum judgment interval, then the robot is determined to be in a stationary state.

[0099] In this embodiment, this step integrates UWB precise positioning and visual verification positioning to construct a relative position representation of the robot. The system obtains the current UWB tag coordinates of the target robot from the spatiotemporal database in step S110 (there is a systematic error within ±0.3 meters under normal working conditions, and may produce decimeter-to-meter jumps under non-line-of-sight (NLOS) or occlusion conditions). It searches the static landmark library for 2-3 landmarks that are closest to the UWB position and whose Euclidean distance is less than the distance threshold (usually set to 3 meters, and the search range is narrowed based on UWB centimeter-level accuracy to ensure the accuracy of landmark association) as reference points. For each reference landmark, the difference between its world coordinates and the robot's UWB coordinates is calculated to form a preliminary relative vector. To improve accuracy and suppress UWB multipath error, the system further utilizes the top view in step S141: by image recognition, the pixel position of the robot in the top view is located (e.g., the robot's center point is located at image coordinates (320, 240)), and combined with the top view pixel-world coordinate mapping relationship, the visual relative vector of the robot relative to the landmark is calculated. The UWB vector and the visual vector are fused with a weight of 0.7:0.3 (under normal signal quality; if a UWB coordinate jump is detected exceeding the threshold or entering a base station coverage blind zone, the weight is dynamically adjusted to 0.2:0.8 with vision as the dominant factor), resulting in the final relative position vector. This vector is then concatenated to form the robot's position fingerprint (e.g., [Δx=3.21m, Δy=1.15m to landmark A; Δx=-1.08m, Δy=2.33m to landmark B]). This fingerprint is used in step S150 to correct the robot's pose: when the UWB changes due to occlusion, multipath effects, or temporary coordinate failure, the system uses the fingerprint to infer the robot's correct world coordinates, achieving continuous and reliable centimeter-level positioning. For determining a stationary state, the system caches the location fingerprint sequence within the past 5 seconds. If the vector difference of 10 consecutive fingerprints is less than the tolerance range (e.g., 0.05 meters) and the time interval exceeds the minimum judgment interval (e.g., 3 seconds), the robot is determined to be in a true stationary state (effectively distinguishing between UWB signal jitter and real movement). This determination result is fed back to step S110 to update the robot's status flag and used in the task allocation logic in step S160.

[0100] Specifically, the weights for normal operating conditions are: w_UWB=0.7, w_vision=0.3; Abnormal operating condition weights: w_UWB=0.2, w_vision=0.8; Anomaly detection criteria: UWB coordinate jump threshold: Δp_UWB>0.5 meters (deviation from predicted position); Or UWB signal loss (entering a base station coverage blind spot); Parameters for determining stillness: Cache sequence length: N_cache=10 (corresponding to 5 seconds @ 2Hz sampling); Vector difference tolerance: Δf_th = 0.05 meters; Minimum decision interval: T_min = 3.0 seconds.

[0101] S150. Based on the obstacle spatiotemporal trajectory prediction results, the robot position fingerprint, the multi-robot collaborative perception results, and the corresponding trusted weights of each robot, the obstacle avoidance strategy on the planned path is dynamically determined and the path is replanned to generate a collaborative scheduling instruction.

[0102] This step is the decision-making center of the entire obstacle avoidance system. It receives the outputs of all preceding perception and prediction modules, transforming abstract obstacle trajectory predictions and robot state representations into concrete, executable multi-robot collaborative action commands. Its core logic is "assess first, then decide": First, an individualized conflict risk assessment is performed on each robot, quantifying the degree of danger of its current planned path and the dynamic environment; then, based on the risk level, hierarchical decision-making is implemented, triggering differentiated response strategies for different types of obstacles (static, dynamic, temporary). During the decision-making process, perception data from robots with high confidence weights account for a larger proportion in conflict calculations, ensuring that decisions are based on the most reliable input; while position fingerprints are used to suppress drift and jump errors in UWB positioning caused by non-line-of-sight (NLOS), multipath effects, or electromagnetic interference, ensuring that the robot makes safe decisions based on accurate relative position cognition.

[0103] In this embodiment, the collaborative scheduling instruction refers to a structured control instruction package, which includes the target robot ID, instruction type (deceleration / replanning / waiting / detour), instruction parameters (target speed, sequence of alternative path points, safe waiting distance), execution priority (urgent / normal / recommended), and decision confidence score. This instruction package is broadcast to the target robot through the distributed communication network of the robot cluster, driving step S160 to execute specific actions.

[0104] In one embodiment, step S150 described above may include steps S151 to S152.

[0105] S151. Based on the obstacle spatiotemporal trajectory prediction results and the robot position fingerprint, combined with the multi-robot collaborative perception results and the corresponding credible weights of each robot, a conflict risk assessment is performed on each mobile robot, including: calculating the spatiotemporal overlap between the planned path of each mobile robot and all obstacle prediction risk areas, the estimated collision time and collision probability, and quantifying and generating a conflict risk level.

[0106] In this embodiment, this step implements individualized hazard measurement, generating a real-time safety situation assessment for each robot. The system loads the obstacle spatiotemporal trajectory prediction results output in step S130 (including the predicted trajectory point sequence and risk area polygon), the robot position fingerprint provided in step S140 (used to correct the robot's own localization), the multi-robot collaborative perception results in step S125 (supplementing the latest state of the current obstacle), and the trust weights of each robot determined in step S123, and starts the conflict risk assessment engine.

[0107] Spatiotemporal overlap calculation: For each robot, its planned path (a sequence of polyline segments generated by the global path planner) is geometrically intersected with the predicted risk areas of each obstacle at all future time points. If a path segment intersects with any risk area polygon, the path segment is marked as an overlapping segment, and the proportion of the overlapping length to the total path length is calculated to obtain the spatial overlap (0-1). Simultaneously, the timestamp of the overlap is calculated to obtain the temporal overlap (1 indicates complete synchronization). The two are multiplied to obtain the comprehensive spatiotemporal overlap; a value greater than 0.3 indicates a potential conflict.

[0108] Estimated Time-to-Collision (TTC): For each overlapping segment, calculate the time difference between the time it takes for the robot to reach the overlapping area at its current speed and the time it takes for the predicted trajectory of the obstacle to reach the same area. If the TTC is less than the safety threshold (usually set to 3 seconds), it indicates that the reaction time is tight and the risk level is increased. The smaller the TTC, the higher the urgency.

[0109] Collision probability calculation: Quantitative assessment considering prediction uncertainty. Based on the positional standard deviation of the risk area in step S135, a collision probability model is established. Assuming the robot path is a defined line and the future positions of obstacles follow a Gaussian distribution centered at the prediction point, the overlap integral of the robot profile and obstacle profile within the risk area is calculated to obtain the collision probability in the 0-1 interval. Confidential weights play a role here: obstacle risk areas generated by the perception data of high-weight robots have a higher weight in the probability calculation, while the data contribution of low-weight robots is suppressed, avoiding misjudgments.

[0110] Based on the above three indicators, the system uses a weighted scoring method to quantify conflict risk into three levels: low conflict (overall score < 0.3), medium risk (0.3-0.7), and high risk (> 0.7). This risk level directly determines the strength of the obstacle avoidance strategy adopted in step S152. For example, high-risk robots will be prohibited from entering the new task area, medium-risk robots will trigger speed limits, and low-conflict robots will maintain normal passage.

[0111] Specifically, the spatiotemporal overlap threshold is: O_th = 0.3; Safe collision time threshold: TTC_th = 3.0 seconds; Collision probability model: Gaussian distribution, 95% confidence level; Risk level classification: Low conflict: R < 0.3; Medium risk: 0.3 ≤ R ≤ 0.7; High risk: R>0.7.

[0112] S152. Perform hierarchical decision-making; wherein, when a static obstacle is detected and its continuous tracking exceeds a set number of frames using the spatiotemporal database, it is marked as a fixed obstacle and the global environment map is updated, triggering a detour strategy in the path planning layer; when the predicted risk area of ​​a dynamic obstacle overlaps with the robot's planned path in spatiotemporal terms and the collision time is less than a safety threshold, the probability of path conflict is calculated. If the probability exceeds the first-level threshold, a cooperative deceleration command is sent to the relevant robot. If it exceeds the second-level threshold, a distributed path replanning request is triggered; when a temporary obstruction is detected that persists within a set number of tracking frames but has no fixed features, it is marked as a pause and yield area, and the robot is controlled to wait at a safe distance or initiate a local detour to form a cooperative scheduling command.

[0113] This step, based on the risk assessment results of S151, implements refined decision-making through classification and grading, and generates multiple types of collaborative scheduling instructions.

[0114] Static obstacle handling: For obstacles classified as static in step S125 (such as walls and fixed equipment), the spatiotemporal database from step S110 is queried to verify whether they have persisted and remained in a stable position for the past 50 frames (approximately 5 seconds). If the verification passes, they are marked as fixed obstacles and inserted into the static layer of the global environment map, replacing the original occupied area. Upon receiving the update, the path planning layer immediately triggers a detour strategy: for high-risk robots, the global path is recalculated to completely avoid the area; for medium- and low-risk robots, a detour buffer distance (e.g., 0.5 meters) is added to the local path. This mechanism ensures continuous evolution of the environment map and reduces the overhead of repeated detection.

[0115] Dynamic Obstacle Response: For dynamic targets such as pedestrians and mobile robots, when the collision probability calculated by S151 exceeds the first-level threshold (e.g., 0.4), the system broadcasts a cooperative deceleration command to the affected robots. The command parameters include the target speed (reducing to 50% of the current speed) and the deceleration distance (gradual deceleration 1 meter in advance). This strategy aims to gain more reaction time and avoid sudden braking. If the collision probability exceeds a higher second-level threshold (e.g., 0.7), it indicates that the current path cannot be safely traversed, and the system triggers a distributed path replanning request: the affected robots suspend execution of the original path, start a local RRT or A algorithm, and search for a new path in the real-time updated risk map. The new path must not overlap with any predicted obstacle risk areas and its total length must not increase by more than 30%. After replanning is completed, the robot switches to the new path, and the original path is discarded. This hierarchical response mechanism strikes a balance between safety and efficiency.

[0116] Temporary Obstacle Handling: For temporary obstructions such as stacked boxes and temporarily parked carts, the system verifies that they persist for 20 frames but have unstable characteristics (no fixed shape), marking their area as a pause and yield zone. For robots about to enter this zone, the instructions are divided into two types: If the zone width is less than the robot's detour radius, a wait instruction is sent, controlling the robot to stop at a safe distance (e.g., 1.5 meters), waiting for the obstruction to disappear (confirmed through continuous monitoring) before resuming passage; if the zone width allows, a local detour instruction is sent, inserting a detour waypoint into the local path, and detouring at a low speed of 0.3 meters per second. This strategy avoids frequent global path replanning due to overreaction, maintaining system smoothness.

[0117] All decisions are ultimately encapsulated into collaborative scheduling instructions, which are broadcast to the target robot via ROS topics or a custom UDP protocol. The instructions include: timestamp, instruction ID, target robot ID (single or multiple), instruction type (STATIC_AVOID / DYNAMIC_SLOWDOWN / DYNAMIC_REPLAN / TEMP_WAIT / TEMP_DETOUR), instruction parameters (speed value, waypoint list, waiting area coordinates), and decision confidence (calculated based on the conflict probability in S151 and the confidence weight in S123), ensuring that step S160 is executed accurately. High-frequency instructions (such as deceleration) are updated once per second, while low-frequency instructions (such as replanning) remain active after triggering until completion.

[0118] Specifically, the static obstacle verification frame count is: N_verify = 50 frames (5 seconds @ 10fps). Dynamic obstacle response threshold: Level 1 threshold (cooperative deceleration): P1 = 0.4; Second-level threshold (path replanning): P2 = 0.7; Temporary occlusion verification frame count: N_temp_verify=20 frames (2 seconds @ 10fps); Detour buffer distance: d_buffer = 0.5 meters; Deceleration parameters: Target speed ratio: η = 50%; Deceleration distance: d_slow = 1.0 meter; Replanning constraint: Path length increase <30%; Safe waiting distance: d_wait = 1.5 meters; Local detour speed: v_detour = 0.3 m / s.

[0119] S160. Based on the cooperative scheduling instructions, and combined with the robot's current state and environmental characteristics retrieved from the spatiotemporal database, perform local obstacle avoidance behavior control or global task reallocation, and generate a structured execution log.

[0120] This step serves as the execution and recording layer of the entire robot dynamic obstacle avoidance path planning method. It receives the collaborative scheduling instructions generated in step S150, transforms them into specific actions of the robot itself, and digitally archives the entire process. The system retrieves the robot's state parameters (such as battery level, load, and mobility) and environmental context information (such as static maps and task priorities) required for executing instructions in real time from the dynamic environment spatiotemporal database constructed in step S110. Consensus is reached among the robot clusters through a distributed negotiation mechanism, achieving balanced task load and minimized risk. At the behavior control level, this step abandons a single obstacle avoidance strategy and instead implements differentiated action orchestration based on the obstacle types (static / dynamic / temporary) classified in step S125, ensuring accurate and efficient responses. Finally, all execution details are encapsulated into a structured execution log, providing not only traceability for current decisions but also incremental training data for subsequent model optimization (step S170), forming a closed-loop learning process.

[0121] In this embodiment, the structured execution log refers to a digital record file conforming to the JSON Schema specification, containing all key elements within each decision cycle (usually 1 second): timestamp, obstacle ID and type that triggered the decision, list of robot IDs participating in the collaboration and the trust weight of each robot, scheduling decision type adopted (such as deceleration / replanning / waiting), decision confidence score, execution result status (success / failure / timeout), and multi-robot data association matrix. This log serves as the source of incremental training data for model fine-tuning in step S170.

[0122] In one embodiment, step S160 described above may include steps S161 to S163. S161. Based on the collaborative scheduling instruction, retrieve the robot's current battery level, load weight, and body motion status data from the dynamic environment spatiotemporal database, and use a distributed negotiation mechanism to dynamically redistribute tasks, prioritizing the transfer of high-risk area tasks to low-conflict mobile robots; wherein, low-conflict mobile robots are those whose current planned path or predicted trajectory does not significantly overlap with the predicted risk area of ​​dynamic obstacles, whose collision probability is lower than a preset safety threshold, and which are not in a paused yielding or emergency braking state, based on obstacle spatiotemporal trajectory prediction results and conflict risk assessment.

[0123] In this embodiment, this step achieves decentralized task load balancing and risk avoidance. When step S150 determines that a robot faces a high-risk conflict (e.g., a collision probability exceeding 0.7) and triggers a path replanning request, the system not only requires the robot to adjust its path but also initiates a dynamic task reallocation process. Specifically, the system retrieves the real-time status of all robots in the cluster from the dynamic environment spatiotemporal database in step S110: current battery level (must be above 20% to support additional tasks), load weight (must be below 80% of the rated load to retain maneuverability), body motion status data (maximum speed, maximum acceleration), and current task list. Through a distributed negotiation mechanism (such as the DDS protocol based on ROS2 or a blockchain smart contract), each robot node broadcasts its own status and available capacity, forming a temporary task auction market. High-risk robots decompose their non-urgent tasks (such as inspection point visits and goods delivery) into sub-task units and initiate task transfer offers to low-conflict mobile robots in the network.

[0124] The determination of low-conflict mobile robots is strictly based on the conflict risk level assessment results output in step S151: the geometric intersection area between its planned path and the predicted risk areas of all obstacles is less than 0.1 square meters (no significant spatiotemporal overlap), the estimated collision time is always greater than 5 seconds and the collision probability is less than 0.2 (below the preset safety threshold), and its state machine is not in WAIT (pause and yield) or EMERGENCY_STOP (emergency braking) mode (not in pause and yield or emergency braking state). Such robots are considered nodes with sufficient safety margin and have the ability to undertake additional tasks. During the negotiation process, the system prioritizes the low-conflict robot that is closest to the task starting point, has the most abundant battery, and has the highest trust weight as the task receiver. Once a consensus is reached, the task metadata (target coordinates, deadline, priority) is transferred from the high-risk robot to the receiver, and the global task topology map is updated synchronously to ensure that the overall operation efficiency does not decrease due to local risks. This mechanism effectively avoids the single-point bottleneck and communication delay of the centralized scheduling center and realizes the self-organized risk avoidance of the robot cluster.

[0125] Specifically, the power threshold is: E_min = 20%; Load threshold: L_max = 80% (rated load); Low-collision determination parameters: Spatiotemporal overlap area: A_overlap < 0.1 square meters; Estimated time to collision: TTC > 5.0 seconds; Collision probability: P_collision<0.2.

[0126] S162. For obstacle avoidance behavior control of each mobile robot, a differentiated strategy is executed according to the type of obstacle; wherein, the differentiated strategy includes performing geometrical avoidance for static obstacles; performing speed matching and lateral avoidance for dynamic obstacles; and performing a waiting, detection, and re-passing strategy for temporary obstructions.

[0127] In this embodiment, this step refines the scheduling instructions in S150 into low-level control signals for the robot body, reflecting the refined control concept of classification-based policy implementation. Based on the obstacle type labels identified in step S125, the system calls the corresponding strategy module from the control algorithm library: Static obstacles (such as walls, fixed shelves): Perform geometric avoidance maneuvers. The system extracts the avoidance direction (left / right) and buffer distance (typically 0.5 meters) from the avoidance strategy received in step S150. An intermediate waypoint is inserted between the robot's current position and the target point. This waypoint is offset along the tangent direction of the obstacle's edge, allowing the robot to bypass the obstacle in a smooth arc. The control employs a pure tracking algorithm to track this waypoint, ensuring the robot completes the avoidance with smooth motion not exceeding its maximum acceleration, avoiding sharp turns.

[0128] Dynamic obstacles (such as pedestrians and mobile robots): Speed ​​matching and lateral avoidance are performed. When a cooperative deceleration command is sent in step S150, the robot controller reduces the target speed to the commanded value and uses a PID speed controller to achieve smooth deceleration. If the probability of conflict is high, the robot initiates lateral avoidance: while maintaining a roughly unchanged forward direction, it uses a local path planner to generate a lateral offset trajectory, moving 0.5 meters at a lateral speed of 0.3 m / s to avoid the predicted risk area before returning to the original path. This strategy reduces the relative speed difference through speed synchronization and creates a safe lateral distance through lateral movement, significantly reducing the risk of collision.

[0129] Temporary obstructions (such as temporarily stacked goods): A wait-detect-re-pass strategy is implemented. After receiving the wait instruction in step S150, the robot stops at a safe distance of 1.5 meters from the obstruction and initiates detection mode: taking an image every second and using the perception model from step S120 to determine if the obstruction has disappeared. If the obstruction is not cleared after a preset time limit (e.g., 30 seconds), it is considered a static obstacle, triggering task reassignment in S161; if the obstruction is detected to have been removed within the time limit, the robot automatically switches to the re-pass strategy, cautiously passing through the original area at a low speed of 0.5 m / s, and then resuming full speed after passing. This strategy strikes a balance between flexibility and safety, preventing the robot from getting stuck in an infinite wait.

[0130] Specifically, lateral avoidance parameters: Lateral velocity: v_lat = 0.3 m / s; Lateral displacement: d_lat = 0.5 meters; Wait-probe-re-pass parameters: Detection period: T_detect = 1.0 seconds; Preset timeout: T_timeout = 30 seconds; Resume speed: v_resume = 0.5 m / s.

[0131] S163. Create a structured execution log, which includes obstacle ID, type, obstacle spatiotemporal trajectory prediction results, a list of mobile robot IDs participating in the collaboration, the trust weight of each robot, scheduling decision type, spatiotemporal stamp, and confidence score.

[0132] In this embodiment, this step achieves digital archiving and traceability of the decision-making process. Whenever a collaborative scheduling instruction is generated in step S150 and executed by steps S161 and S162, the system immediately creates a structured execution log record. The log fields strictly adhere to a predefined schema: Obstacle ID and Type: Directly reference the unique identifier and classification label (static / dynamic / temporary) assigned to the obstacle in step S125 to ensure that the source of the problem is traceable.

[0133] Obstacle spatiotemporal trajectory prediction results: The predicted trajectory point sequence and risk area polygon coordinates generated in step S130 are fully recorded to provide a basis for prediction in post-analysis and facilitate the evaluation of prediction accuracy.

[0134] List of mobile robot IDs participating in the collaboration: Records all robot IDs involved in the S161 negotiation process, including high-risk task givers and low-conflict task receivers, reflecting the level of participation in the cluster collaboration.

[0135] Trust weights for each robot: The trust weights for each robot calculated in step S123 are replicated to reflect the data reliability of each node within the decision-making cycle, providing a historical baseline for subsequent weight optimization.

[0136] Scheduling decision type: an enumerated value that records the specific strategy adopted by S152, such as STATIC_AVOID, DYNAMIC_SLOWDOWN, DYNAMIC_REPLAN, TEMP_WAIT, etc., and is used to calculate the triggering frequency and success rate of different strategies.

[0137] Spacetime stamp: Accurately records the time (Unix timestamp, millisecond level) and location (robot world coordinates) of the decision trigger, supporting post-event playback and hotspot area analysis in the spacetime dimension.

[0138] Confidence score: Calculated by combining the collision probability in S151 and the robot confidence weight in S123. The range of 0-1 represents the confidence level of the decision. Log entries with a confidence score below 0.5 will be marked as "requires review" and will be given priority for model optimization in S170.

[0139] The logs are stored in JSON format on a distributed file system (such as HDFS), aggregated into a single file every hour, and synchronously uploaded to the cloud. The log data is not only used for security auditing and troubleshooting, but more importantly, it serves as incremental training data for fine-tuning the model in step S170: by comparing the decision execution results (whether obstacle avoidance was successful) with the predicted labels, the trajectory prediction model in step S130 and the threshold parameters in step S150 are optimized in reverse, forming a continuous learning loop that allows the system performance to continuously improve over time.

[0140] In another embodiment, the method further includes the following after step S160: If the conflict risk determined in step S150 is higher than the preset safety threshold, or if the execution log generated in step S160 contains scheduling instructions that were not successfully executed, then the disputed area is located on the dynamic environment situation map and output to the remote monitoring terminal; in response to the manual review result information input from the monitoring terminal, the corresponding scene data and the judgment result are associated to construct incremental training data; the incremental training data is periodically used to fine-tune the deep learning model and prediction model in steps S120, S130 and S140 to optimize the credible weight allocation strategy in step S120, the trajectory prediction accuracy in step S130 and the static landmark recognition accuracy in step S140.

[0141] This step constitutes a closed-loop learning and continuous optimization mechanism for the entire robot dynamic obstacle avoidance path planning method. It aims to automatically capture system misjudgments or performance bottlenecks through human-machine collaboration, driving iterative model upgrades. When the autonomous decision-making system encounters extremely complex scenarios or algorithmic limitations, it triggers human expert intervention, transforming expert experience into structured training data to achieve a virtuous cycle of "running-making mistakes-learning-improvement," ensuring continuous improvement in system performance over deployment time. This mechanism not only enhances the system's robustness in long-tail scenarios but also precisely addresses core pain points in multi-robot collaboration, such as data reliability, motion uncertainty, and positioning drift, through targeted optimization of three core modules: reliable weights, trajectory prediction, and landmark recognition.

[0142] In one embodiment, step S170 described above may include steps S171 to S173.

[0143] S171. When the conflict risk level generated in step S151 is high risk (conflict probability > 0.7), or when the structured execution log created in step S163 records an entry with the scheduling decision execution status as "FAILED" (unsuccessful execution), the system locates the disputed area on the dynamic environment situation map and outputs it to the remote monitoring terminal.

[0144] This step establishes an automated problem detection and alert mechanism. A high-risk determination means that the collision probability calculated in S151 has exceeded the system's safety tolerance limit, indicating that the current trajectory prediction or conflict assessment may have serious deviations, and the robot is on the verge of danger even though it has not actually collided. Execution failure refers to the robot's failure to complete the action as instructed in S162, for example: receiving a replanning instruction but still unable to find a feasible path due to a sudden change in the environment, or failing to respond to a deceleration instruction due to a motor malfunction. Both of these situations suggest deficiencies in the algorithm model.

[0145] The system then activates the disputed area localization: extracting the predicted risk area polygon coordinates at high-risk moments from the obstacle spatiotemporal trajectory prediction results output by S130, extracting the distribution of static landmarks around the robot at that moment from S140, and extracting all robot pose and sensor data from S110. This information is then overlaid onto a dynamic environmental situation map—a real-time visualization layer based on a GIS framework, containing semantic information such as road geometry, obstacle locations, robot trajectories, and static landmarks. The disputed area is highlighted in red on the situation map, accompanied by key metadata: the timestamp of the dispute, the robot ID involved, the conflict probability value, and the type of failure command. The entire situation map is pushed to the web interface of the remote monitoring center via 4G / 5G network or Wi-Fi. Monitoring personnel can view it in real time through a browser and click on the markers to access multi-view video playback, sensor data curves, and decision logic chains of the disputed area, enabling rapid problem diagnosis.

[0146] S172. Respond to the manual review result information input from the monitoring terminal, and perform multimodal association between the review result and the corresponding scene data and judgment result to construct incremental training data.

[0147] This step structures expert experience into knowledge. Manual review involves monitoring personnel evaluating the system's decisions on the monitoring interface based on experience and video evidence, selecting a review result label: TRUE_POSITIVE (true positive, the system correctly identified a real hazard), FALSE_POSITIVE (false positive, the system misjudged a safe area as dangerous), TRUE_NEGATIVE (true negative, the system correctly ignored a safe scenario), and FALSE_NEGATIVE (false negative, the system missed an actual hazard). For example, if the system misinterprets a water stain on the ground as a dynamic obstacle due to glare, triggering emergency deceleration, the monitoring personnel should mark it as FALSE_POSITIVE.

[0148] Upon receiving the review label, the system automatically initiates multimodal association: binding the label to the complete scene data at the moment of the dispute, including raw perception data (RGB image and depth image in S111), intermediate perception results (obstacle feature information and confidence weights in S125), prediction output (predicted risk area coordinates in S135 and kinematic model parameters in S133), environmental context (static landmark image patches in S142 and robot state data in S110), and decision logs (scheduling instructions in S163 and conflict probability calculation process in S151). All data is timestamped, packaged into an incremental training sample, and stored in JSON format in the training dataset of a distributed storage system. Each sample also includes a decision link fingerprint, recording which module caused the misjudgment (e.g., unreasonable weight allocation in S123 or excessive speed estimation in S133), providing guidance for subsequent targeted optimization.

[0149] S173. Periodically use the incremental training data to fine-tune the deep learning model and prediction model in steps S120, S130 and S140, and optimize the credible weight allocation strategy in step S123, the trajectory prediction accuracy in step S133 and the static landmark recognition accuracy in step S142.

[0150] This step completes the knowledge internalization from data to model. The system automatically triggers a periodic fine-tuning training task every N incremental samples (N is configurable, usually set to 100) or every T hours (usually 24 hours), using an online learning mode to avoid model obsolescence.

[0151] Specifically, the incremental sample trigger threshold is: N_trigger = 100. Time-triggered threshold: T_trigger = 24 hours; Fine-tuning the learning rate: η_fine = 0.0001; A / B testing ratio: 10% (grayscale release); Success rate improvement threshold: ΔS=5% (for full-scale promotion). The optimization strategy for the credible weight allocation in step S123 involves fine-tuning the feature fusion branch in the dual-branch deep learning network (S124). Using samples labeled FALSE_POSITIVE or FALSE_NEGATIVE, the fusion weights of semantic features and fine-grained features are adjusted. For example, if a large number of misclassifications are caused by sudden changes in illumination, the model will increase the weights of brightness-invariant features (such as texture and edges) and decrease the dependence on color features. Fine-tuning employs a transfer learning strategy: the bottom convolutional layers of the pre-trained large model are frozen, and only the parameters of the top fully connected layers are updated, with a learning rate of 0.0001 to prevent catastrophic forgetting. After training, the model's feature quality assessment capability under similar illumination scenarios is significantly improved, and the output fused feature vector is more accurate. This makes the correlation coefficient calculated in S123 more sensitive to the quality of real data, and the credible weight allocation more precise.

[0152] Optimize the trajectory prediction accuracy in step S133: Fine-tune the kinematic prediction model. Using missed samples labeled FALSE_NEGATIVE (especially in scenarios with sudden turning due to dynamic obstacles), analyze the deviation between the motion baseline fitted in S133 and the actual trajectory. Using the actual trajectory data as the ground truth, recalibrate the expected velocity calculation method in the model using maximum likelihood estimation: introduce an acceleration decay factor so that the model considers not only historical average velocity but also the obstacle category (pedestrians have higher turning flexibility than forklifts) when making predictions, dynamically adjusting the speed maintenance coefficient. The fine-tuned model reduces prediction errors in complex scenarios such as curves and intersections, and the predicted risk area generated in S135 better reflects the uncertainty of real motion.

[0153] Optimize the accuracy of static landmark recognition in step S142: fine-tune the feature point detection algorithm (e.g., SuperPoint). Use the precise locations of static landmarks manually marked on the monitoring end as supervision signals to compare the landmark coordinate errors extracted in S142. Through few-shot learning, with the encoder frozen, only the detection head in the decoder is updated, incrementally learning the features of special landmarks in the new environment (e.g., corner points of shelf labels in a warehouse). After fine-tuning, the landmark repetition detection rate is improved, the position error is reduced, making the robot position fingerprint constructed in S143 more accurate, thereby improving the positioning reliability of the S150 decision.

[0154] After fine-tuning, the new model was released in a phased rollout using an A / B testing mechanism: it was first deployed to 10% of the robot nodes and run for 24 hours, comparing the success rate of its decisions with that of the unoptimized nodes. If the success rate improved by more than 5%, it was then rolled out to all nodes. The entire S170 process enabled the system to self-evolve, allowing the robot cluster to continuously adapt to environmental changes and task requirements throughout the deployment cycle, significantly reducing the long-tail failure rate.

[0155] The aforementioned robot dynamic obstacle avoidance path planning method constructs a dynamic spatiotemporal database by collecting environmental data from edge computing terminals mounted on multiple mobile robots. It then verifies the validity and reliability weights of cross-terminal data using spectral similarity and correlation coefficients, thus resolving the issue of data reliability discrepancies across multiple robots. This method preprocesses images to identify static and dynamic obstacles, and combines this with cross-robot multi-frame data correlation to predict the future position and trajectory of dynamic obstacles, resulting in accurate spatiotemporal trajectory predictions and improving the accuracy of dynamic obstacle intention prediction. Simultaneously, static landmarks are used to correct positioning errors, improving positioning accuracy and resolving positioning drift issues. Based on the above information and the reliability weights of each robot, the system can make intelligent decisions, dynamically adjust obstacle avoidance strategies and path planning, generate collaborative scheduling instructions, and guide local obstacle avoidance or global task reallocation, ensuring that the robot swarm achieves high-precision, low-latency, continuous, and safe autonomous obstacle avoidance and path planning in complex dynamic environments. This comprehensive method fully leverages the advantages of multi-robot collaborative perception, achieving reliable data fusion and intelligent decision-making.

[0156] In another embodiment, the aforementioned UWB content can employ a UWB+LiDAR fusion scheme, in which case the overall steps are as follows: The system acquires images, UWB tag coordinates, LiDAR point cloud data, timestamps, and body motion state data collected by edge computing terminals mounted on multiple mobile robots, and constructs a dynamic environment spatiotemporal database based on the collected data. The image is preprocessed to identify and classify static obstacles, dynamic obstacles, and temporary occlusion targets. The obstacle feature information, confidence score, and motion vector are output to obtain the multi-robot collaborative perception results. Furthermore, the cross-terminal data in the dynamic environment spatiotemporal database is subjected to interference removal and validity verification using spectral similarity calculation and correlation coefficient verification methods to obtain the corresponding confidence weights for each robot. Based on the multi-robot collaborative perception results, dynamic obstacles close to the robot's planned path are selected and their future positions are predicted. The motion trajectory of the dynamic obstacles is constructed by associating cross-robot, continuous multi-frame data, and calibrated in a unified world coordinate system to generate a predicted risk area, thus forming an obstacle spatiotemporal trajectory prediction result. The obstacle spatiotemporal trajectory prediction result is associated with the timestamps in the spatiotemporal database, UWB tag coordinates, and LiDAR pose estimation results. In the preprocessed image, a feature point detection algorithm is used to identify and extract fixed static landmarks. The relative position vector of the target robot with respect to the static landmarks is calculated based on the fusion positioning coordinates of UWB and LiDAR. The robot position fingerprint is constructed to suppress UWB non-line-of-sight errors, correct positioning errors, and verify the stationary state. Based on the obstacle spatiotemporal trajectory prediction results, the robot position fingerprint, the multi-robot collaborative perception results, and the corresponding trusted weights of each robot, the obstacle avoidance strategy on the planned path is dynamically determined and the path is replanned to generate collaborative scheduling instructions. Based on the collaborative scheduling instructions, and combined with the robot's current state and environmental characteristics retrieved from the spatiotemporal database, local obstacle avoidance behavior control or global task reallocation is performed, and a structured execution log is generated.

[0157] Furthermore, the acquisition of images, UWB tag coordinates, LiDAR point cloud data, timestamps, and body motion state data collected by edge computing terminals mounted on multiple mobile robots, and the construction of a dynamic environment spatiotemporal database based on the collected data, includes: The edge computing devices of each robot collect RGB images, depth images, UWB tag coordinates (calculated based on TOF or TDOA ranging algorithms), LiDAR point cloud data, timestamps, and the robot's linear velocity, angular velocity, heading angle data, current battery level, load weight, and motion status data in real time to obtain detection data; among which, the robot's motion status data includes maximum speed, maximum acceleration, and braking distance. A UWB base station network matching algorithm combined with LiDAR point cloud inter-frame registration (ICP / NDT) is used to perform spatiotemporal alignment and correlation of detection data of different robots passing through the same location, forming a dynamic environment spatiotemporal database in a unified coordinate system with UWB base stations as the global reference and LiDAR SLAM as the local refinement. The dynamic environment spatiotemporal database includes obstacle positions, static landmark coordinates, robot pose, power load status, body motion status data, and data confidence scores.

[0158] The image is preprocessed to identify and classify static obstacles, dynamic obstacles, and temporary occlusions. Obstacle feature information, confidence scores, and motion vectors are output to obtain multi-robot collaborative perception results. Furthermore, spectral similarity calculation and correlation coefficient verification methods are used to remove interference and verify the validity of cross-terminal data in the dynamic environment spatiotemporal database, obtaining the corresponding trust weights for each robot, including: The images are subjected to denoising and enhancement normalization preprocessing to eliminate the effects of sensor noise and uneven illumination, resulting in a standardized image sequence that conforms to a uniform resolution and color space. For the detection data of different mobile robots on the same obstacle within the same time window in the dynamic environment spatiotemporal database, the similarity metric of the spectral response curve is calculated. Outlier data with similarity below a preset threshold are identified as interference signals and removed. Only valid detection data with spectral consistency are retained to obtain valid data. For the valid data, the ambient illumination value and the average gray value of the image are extracted and the Pearson correlation coefficient between them is calculated. For a single robot, the Pearson correlation coefficient measures the internal perception consistency. For a multi-robot cluster, the distribution of the Pearson correlation coefficients of each robot is compared. When the coefficient of a robot deviates from the group mean by more than a set standard deviation, the perception credibility is determined to be abnormal and the data weight is reduced to obtain the credibility weight corresponding to each robot. Meanwhile, based on the UWB ranging signal-to-noise ratio (SNR) and LiDAR point cloud registration residuals, a data quality score is calculated: when the UWB signal SNR is lower than the threshold or the LiDAR registration score is abnormal, the robot's credibility weight is reduced. The pre-trained large model is called to extract semantic verification features from the standardized image sequence. At the same time, a dual-branch deep learning network is used to detect fine-grained element features that conform to the pixel size range. The two types of features are mapped to a unified high-dimensional feature space and then weighted and fused to generate a fusion feature vector that comprehensively represents the data quality. Based on the effective data, fused feature vectors, and the confidence weights of each robot, static obstacles, dynamic obstacles, and temporary occlusion targets are identified and classified. Motion vectors are calculated, and multi-robot collaborative perception results containing obstacle feature information, confidence scores, and motion vectors are generated.

[0159] Based on the multi-robot collaborative perception results, dynamic obstacles close to the robot's planned path are selected and their future positions are predicted. The motion trajectories of these dynamic obstacles are constructed by associating cross-robot, continuous multi-frame data, and calibrated in a unified world coordinate system to generate predicted risk areas, thus forming the spatiotemporal trajectory prediction results for obstacles. This includes: Based on the multi-robot collaborative perception results, target dynamic obstacles close to the robot's planned path are identified and selected. Spatial distance is calculated based on the UWB tag coordinates of each robot. The detection data of the same obstacle under the multi-robot perspective are associated to obtain the target obstacle after cross-view association. The analysis of the target obstacles in a continuous time-standardized image sequence includes: applying a multi-target tracking algorithm to establish a motion trajectory for each obstacle, and transforming the trajectory to a unified world coordinate system constructed by fusing the UWB base station coordinate system and LiDAR SLAM to obtain spatiotemporally standardized trajectory data; By analyzing the spatiotemporally standardized trajectory data of multiple obstacles in the same scene, calculating the overall motion vector, fitting the obstacle motion baseline and expected velocity, a kinematic prediction model is obtained. Based on the kinematic prediction model, the world coordinates of the predicted point at the future time are determined by extrapolating the average movement distance along the motion baseline, with the current obstacle position as the prediction starting point. Based on the predicted point's world coordinates, a predicted risk region is created in a unified world coordinate system with the predicted point as the center. The vertical axis is a number of times the position standard deviation, and the horizontal axis is a number of times the obstacle size. The predicted risk region is then transformed back to each robot's local coordinate system to obtain the obstacle's spatiotemporal trajectory prediction result.

[0160] In the preprocessed image, a feature point detection algorithm is used to identify and extract fixed static landmarks. Based on the fusion of UWB and LiDAR positioning coordinates, the relative position vector of the target robot with respect to the static landmarks is calculated to construct a robot position fingerprint. This fingerprint is used to suppress UWB non-line-of-sight errors, correct positioning errors, and verify the stationary state. The image is converted into a distortion-free top view using a camera calibration matrix; In the top view, a feature point detection algorithm is used to identify and extract fixed static landmarks, including wall corners, column corners, and fixed equipment corners; at the same time, based on LiDAR point cloud data, RANSAC plane extraction or corner detection algorithms are used to identify three-dimensional static landmarks. Based on the fusion coordinates of coarse UWB coordinate positioning and fine LiDAR positioning, the relative position vector of the target robot with respect to static landmarks at the required distance is calculated as the robot's position fingerprint. When the UWB signal is blocked and causes a jump, the position fingerprint continuity is maintained by prioritizing the matching results between LiDAR odometry and static landmarks. When the difference in position fingerprints at different times is within the set tolerance range and the time interval exceeds the preset minimum judgment interval, the robot is determined to be in a stationary state.

[0161] The method involves combining the obstacle spatiotemporal trajectory prediction results with the robot's position fingerprint, multi-robot collaborative perception results, and the corresponding trust weights of each robot to dynamically determine the obstacle avoidance strategy and replan the path on the planned path, generating collaborative scheduling instructions, including: Based on the obstacle spatiotemporal trajectory prediction results, the robot position fingerprint, the multi-robot collaborative perception results, and the corresponding trust weights of each robot, a conflict risk assessment is performed on each mobile robot, including: calculating the spatiotemporal overlap between the planned path of each mobile robot and all obstacle prediction risk areas, the estimated collision time and collision probability, and quantifying and generating a conflict risk level. Layered decision-making is performed. When a static obstacle is detected and its continuous tracking exceeds a set number of frames using the spatiotemporal database, it is marked as a fixed obstacle, and the LiDAR grid map or point cloud map is updated, triggering a detour strategy in the path planning layer. When the predicted risk area of ​​a dynamic obstacle overlaps with the robot's planned path in both time and space and the collision time is less than a safety threshold, the probability of path conflict is calculated. If the probability exceeds the first-level threshold, a cooperative deceleration command is sent to the relevant robot. If it exceeds the second-level threshold, a distributed path replanning request is triggered. When a temporary occlusion is detected that persists within a set number of tracking frames but has no fixed features, it is marked as a pause and yield area. The robot is controlled to wait at a safe distance or initiate a local detour to form a cooperative scheduling command.

[0162] In this embodiment, UWB provides global consistency constraints (unified coordinate system for multiple machines, time synchronization), while LiDAR provides local high-precision positioning and environmental perception (SLAM, static obstacle detection). The two are fused through Kalman filtering or factor graph optimization, and automatically switch to LiDAR-dominated mode when the UWB signal is blocked to ensure system robustness.

[0163] The difference from the previous embodiment is that: The UWB+LiDAR fusion positioning architecture is implemented as follows: A global coordinate reference (a local coordinate system with the base station as the origin) is constructed using a UWB base station network. The UWB tag mounted on the robot calculates the real-time coordinates (accuracy 10-30cm) using TOF or TDOA algorithms. At the same time, LiDAR estimates high-precision local pose and generates a point cloud map using SLAM algorithms (LOAM or Lego-LOAM). The two are tightly coupled and fused through extended Kalman filter (EKF) or factor graph optimization. Under normal operating conditions, UWB mainly constrains the cumulative drift of LiDAR. When the SNR of the UWB signal is detected to be lower than the threshold (indicating NLOS occlusion) or the coordinate jump exceeds the LiDAR prediction uncertainty, the system automatically switches to the LiDAR visual odometry (VO) dominant mode. Historical map feature matching is used to maintain the continuity of positioning, solving the jump problem of pure UWB in multipath environment and the long-term drift problem of pure LiDAR.

[0164] Spatiotemporal database construction and alignment mechanism: Utilizing the inherent nanosecond-level clock synchronization capability of the UWB base station network (through wired synchronization or IEEE 1588 PTP wireless synchronization protocol) to unify the timestamps of multiple robots and eliminate microsecond-level clock deviations; spatial alignment adopts a two-stage registration strategy—first, coarse alignment is performed based on the coordinates of each robot's UWB tag (converting the local ranging solution coordinates to a unified world coordinate system with the main base station as the origin), followed by fine alignment through LiDAR point cloud inter-frame registration (ICP or NDT algorithm), calculating the registration residual as a data quality indicator; finally, the spatiotemporal database stores the fused precise pose (UWB global coordinates + LiDAR local refinement), obstacle 3D point cloud positions, and data confidence scores calculated jointly based on UWB SNR and registration residuals.

[0165] Cross-terminal data verification and trusted weight calculation: Based on the original spectral similarity (FFT frequency domain feature consistency) verification, UWB signal quality assessment (range signal-to-noise ratio SNR, multipath indicator) and LiDAR geometric consistency verification (point cloud registration residual, feature matching inlier rate) are introduced; when a robot's UWB SNR is lower than the NLOS detection threshold (e.g., <30dB) or the LiDAR registration residual exceeds 5cm, it is determined that the positioning data at that moment is occluded or the sensor is abnormal, and its trusted weight is dynamically reduced (reduced by a coefficient of 0.3-0.7); during multi-machine collaborative perception, only the detection data of robots with high trusted weight (>0.8) are retained to participate in obstacle association and trajectory fusion, effectively suppressing false alarms caused by UWB multipath error or LiDAR dynamic object interference.

[0166] Specifically, the UWBSNR threshold is: SNR_th = 30dB; LiDAR registration residual threshold: ε_LiDAR=5cm; Coordinate jump threshold: Δp_jump=50cm (UWB and LiDAR predicted position deviation); Weight reduction factor: γ = 0.3~0.7 (linear mapping); High-confidence weight threshold: w_high=0.8; Spatial correlation threshold (tighter): D_tight = 2.0 meters; Dynamic adjustment of safety margin: UWB signal is good: d_margin=0.5 meters; NLOS status: d_margin = 1.0 meters.

[0167] Cross-robot obstacle association and trajectory prediction: Based on UWB centimeter-level positioning accuracy, a tighter spatial association threshold (typically 2 meters) is set. When multiple robots detect obstacles with a spatial distance less than the threshold and of the same category within a ±0.5-second time window, cross-view association is triggered. The precise 3D geometric contours provided by LiDAR point clouds (through Euclidean clustering or deep learning segmentation) are used to establish cross-robot ID mapping in combination with the Hungarian algorithm or DeepSORT multi-object tracking framework. In a unified world coordinate system (UWB global baseline + LiDAR local refinement), extended Kalman filter (EKF) is used to fuse the observation trajectories of multiple robots. The state vector includes obstacle position (X,Y), velocity (Vx,Vy) and UWB positioning uncertainty covariance matrix. When extrapolating future positions, UWB measurement noise is considered (dynamically adjusting the process noise Q matrix), and elliptical or rectangular predicted risk areas are generated (the vertical axis is the position standard deviation multiple, and the horizontal axis is the obstacle size multiple).

[0168] Location fingerprint construction and static state verification mechanism: Based on the fusion results of UWB coordinates and LiDAR pose estimation, the relative position vector [Δx, Δy] of the target robot with respect to the nearest 2-3 static landmarks (intersection points of wall corners / columns extracted from LiDAR point clouds using the RANSAC algorithm, combined with visual ORB / SIFT feature points) is calculated. When the UWB coordinates change due to NLOS (deviation from LiDAR predicted position > 50cm and a sharp drop in SNR), the system automatically removes the abnormal UWB data and uses LiDAR visual matching (_SCAN context_ or _BoW3D_ descriptor) to maintain fingerprint continuity. Static determination adopts dual threshold verification - not only is the fingerprint vector difference less than the tolerance (0.05m) required for 10 consecutive frames (5 seconds), but also the LiDAR point cloud inter-frame registration score (<2cm displacement) and UWB coordinate stability (<20cm fluctuation) need to be verified to effectively distinguish UWB signal jitter, slipping in place and true static state.

[0169] Dynamic environment map update and obstacle avoidance decision-making: Occupied grid map (OGM) or 3D point cloud map is constructed and updated in real time using LiDAR point cloud. Static obstacles (continuously tracked for more than 30 frames) are fixed to the global map through voxel filtering downsampling and dynamic object removal (based on multi-frame differential or Ray casting). In obstacle avoidance decision-making, the precise relative distance between the robot and obstacles is calculated based on UWB+LiDAR fusion pose (error <10cm). When the predicted risk area of ​​a dynamic obstacle is detected to overlap with the planned path, the safety threshold is dynamically adjusted according to the UWB positioning accuracy (0.5m safety margin is used when the UWB signal is good, and it is extended to 1m in NLOS). Cooperative scheduling instructions containing speed adjustment and alternative path point sequences are generated and broadcast to the target robot for execution via ROS2 DDS or MQTT.

[0170] Figure 2 This is a schematic block diagram of a robot dynamic obstacle avoidance path planning system 300 provided in an embodiment of the present invention. Figure 2 As shown, corresponding to the above-described robot dynamic obstacle avoidance path planning method, the present invention also provides a robot dynamic obstacle avoidance path planning system 300. This robot dynamic obstacle avoidance path planning system 300 includes a unit for executing the above-described robot dynamic obstacle avoidance path planning method, and the system can be configured in a server. For details, please refer to... Figure 2 The robot dynamic obstacle avoidance path planning system 300 includes an acquisition unit 301, a processing unit 302, a prediction unit 303, a fingerprint construction unit 304, an instruction generation unit 305, and an execution unit 306.

[0171] The acquisition unit 301 is used to acquire images, UWB tag coordinates, timestamps, and body motion state data collected by edge computing terminals mounted on multiple mobile robots, and construct a dynamic environment spatiotemporal database based on the acquired data. The processing unit 302 is used to preprocess the images, identify and classify static obstacles, dynamic obstacles, and temporary occlusion targets, output obstacle feature information, confidence scores, and motion vectors, obtain multi-robot collaborative perception results, and use spectral similarity calculation and correlation coefficient verification methods to perform interference removal and validity verification on cross-terminal data in the dynamic environment spatiotemporal database to obtain the corresponding confidence weights for each robot. The prediction unit 303 is used to select dynamic obstacles close to the robot's planned path and predict their future positions based on the multi-robot collaborative perception results, construct the dynamic obstacle motion trajectory by associating cross-robot, continuous multi-frame data, and calibrate in a unified world coordinate system to generate predicted risk areas, forming obstacles. The system includes: a spatiotemporal trajectory prediction result for obstacles; wherein the spatiotemporal trajectory prediction result for obstacles is associated with timestamps in the spatiotemporal database and UWB tag coordinates; a fingerprint construction unit 304, used to identify and extract fixed static landmarks in the preprocessed image using a feature point detection algorithm, calculate the relative position vector of the target robot relative to the static landmarks based on the UWB tag coordinates, and construct a robot position fingerprint to correct positioning errors and verify the stationary state; an instruction generation unit 305, used to dynamically determine and replan the obstacle avoidance strategy on the planned path based on the obstacle spatiotemporal trajectory prediction result and the robot position fingerprint, combined with the multi-robot collaborative perception result and the corresponding confidence weight of each robot, and generate a collaborative scheduling instruction; and an execution unit 306, used to execute local obstacle avoidance behavior control or global task reallocation based on the collaborative scheduling instruction, combined with the robot's current state and environmental features retrieved from the spatiotemporal database, and generate a structured execution log.

[0172] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned robot dynamic obstacle avoidance path planning system 300 and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0173] The aforementioned robot dynamic obstacle avoidance path planning system 300 can be implemented as a computer program, which can, for example... Figure 3 It runs on the computer device shown.

[0174] Please see Figure 3 , Figure 3This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0175] See Figure 3 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0176] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a robot dynamic obstacle avoidance path planning method.

[0177] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0178] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a robot dynamic obstacle avoidance path planning method.

[0179] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0180] The processor 502 is used to run the computer program 5032 stored in the memory to implement all the steps of the robot dynamic obstacle avoidance path planning method.

[0181] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0182] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0183] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform all steps of the robot dynamic obstacle avoidance path planning method.

[0184] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0185] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0186] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0187] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0188] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0189] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for dynamic obstacle avoidance path planning for robots, characterized in that, include: The system acquires images, UWB tag coordinates, timestamps, and body motion state data collected by edge computing terminals mounted on multiple mobile robots, and constructs a dynamic environment spatiotemporal database based on the collected data. The image is preprocessed to identify and classify static obstacles, dynamic obstacles, and temporary occlusion targets. The obstacle feature information, confidence score, and motion vector are output to obtain the multi-robot collaborative perception results. Furthermore, the cross-terminal data in the dynamic environment spatiotemporal database is subjected to interference removal and validity verification using spectral similarity calculation and correlation coefficient verification methods to obtain the corresponding confidence weights for each robot. Based on the multi-robot collaborative perception results, dynamic obstacles close to the robot's planned path are selected and their future positions are predicted. The dynamic obstacle's trajectory is constructed by associating cross-robot, continuous multi-frame data, and calibrated in a unified world coordinate system to generate a predicted risk area, thus forming an obstacle spatiotemporal trajectory prediction result. The obstacle spatiotemporal trajectory prediction result is associated with the timestamps in the spatiotemporal database and the UWB tag coordinates. In the preprocessed image, a feature point detection algorithm is used to identify and extract fixed static landmarks. The relative position vector of the target robot with respect to the static landmarks is calculated based on the UWB label coordinates. The robot position fingerprint is constructed to correct the positioning error and verify the stationary state. Based on the obstacle spatiotemporal trajectory prediction results, the robot position fingerprint, the multi-robot collaborative perception results, and the corresponding trusted weights of each robot, the obstacle avoidance strategy on the planned path is dynamically determined and the path is replanned to generate collaborative scheduling instructions. Based on the collaborative scheduling instructions, and combined with the robot's current state and environmental characteristics retrieved from the spatiotemporal database, local obstacle avoidance behavior control or global task reallocation is performed, and a structured execution log is generated.

2. The robot dynamic obstacle avoidance path planning method according to claim 1, characterized in that, The process involves acquiring images, UWB tag coordinates, timestamps, and body motion state data collected by edge computing terminals mounted on multiple mobile robots, and constructing a dynamic spatiotemporal database based on the collected data, including: The edge computing devices of each robot collect RGB images, depth images, UWB tag coordinates, timestamps, and data on the robot's linear velocity, angular velocity, heading angle, current battery level, load weight, and motion status in real time to obtain detection data; among which, the motion status data includes maximum speed, maximum acceleration, and braking distance. A UWB base station coordinate matching algorithm is used to perform spatiotemporal alignment and correlation of detection data of different robots passing through the same location, forming a dynamic environment spatiotemporal database under a unified coordinate system based on the UWB base station network. The dynamic environment spatiotemporal database includes obstacle positions, static landmark coordinates, robot pose, power load status, body motion status data, and data confidence scores.

3. The robot dynamic obstacle avoidance path planning method according to claim 1, characterized in that, The image is preprocessed to identify and classify static obstacles, dynamic obstacles, and temporary occlusions. Obstacle feature information, confidence scores, and motion vectors are output to obtain multi-robot collaborative perception results. Furthermore, spectral similarity calculation and correlation coefficient verification methods are used to remove interference and verify the validity of cross-terminal data in the dynamic environment spatiotemporal database, obtaining the corresponding trust weights for each robot, including: The images are subjected to denoising and enhancement normalization preprocessing to eliminate the effects of sensor noise and uneven illumination, resulting in a standardized image sequence that conforms to a uniform resolution and color space. For the detection data of different mobile robots on the same obstacle within the same time window in the dynamic environment spatiotemporal database, the similarity metric of the spectral response curve is calculated. Outlier data with similarity below a preset threshold are identified as interference signals and removed. Only valid detection data with spectral consistency are retained to obtain valid data. For the valid data, the ambient illumination value and the average gray value of the image are extracted and the Pearson correlation coefficient between them is calculated. For a single robot, the Pearson correlation coefficient measures the internal perception consistency. For a multi-robot cluster, the distribution of the Pearson correlation coefficients of each robot is compared. When the coefficient of a robot deviates from the group mean by more than a set standard deviation, the perception credibility is determined to be abnormal and the data weight is reduced to obtain the credibility weight corresponding to each robot. The pre-trained large model is called to extract semantic verification features from the standardized image sequence. At the same time, a dual-branch deep learning network is used to detect fine-grained element features that conform to the pixel size range. The two types of features are mapped to a unified high-dimensional feature space and then weighted and fused to generate a fusion feature vector that comprehensively represents the data quality. Based on the effective data, fused feature vectors, and the confidence weights of each robot, static obstacles, dynamic obstacles, and temporary occlusion targets are identified and classified. Motion vectors are calculated, and multi-robot collaborative perception results containing obstacle feature information, confidence scores, and motion vectors are generated.

4. The robot dynamic obstacle avoidance path planning method according to claim 1, characterized in that, Based on the multi-robot collaborative perception results, dynamic obstacles close to the robot's planned path are selected and their future positions are predicted. The motion trajectories of these dynamic obstacles are constructed by associating cross-robot, continuous multi-frame data, and calibrated in a unified world coordinate system to generate predicted risk areas, thus forming the spatiotemporal trajectory prediction results for obstacles. This includes: Based on the multi-robot collaborative perception results, target dynamic obstacles close to the robot's planned path are identified and selected. The detection data of the same obstacle under the perspectives of multiple robots are associated to obtain the target obstacle after cross-perspective association. The analysis of the target obstacles in a continuous time-standardized image sequence includes: applying a multi-target tracking algorithm to establish a motion trajectory for each obstacle, and transforming the trajectory to a unified world coordinate system to obtain spatiotemporally standardized trajectory data; By analyzing the spatiotemporally standardized trajectory data of multiple obstacles in the same scene, calculating the overall motion vector, fitting the obstacle motion baseline and expected velocity, a kinematic prediction model is obtained. Based on the kinematic prediction model, the world coordinates of the predicted point at the future time are determined by extrapolating the average movement distance along the motion baseline, with the current obstacle position as the prediction starting point. Based on the predicted point's world coordinates, a predicted risk region is created in a unified world coordinate system with the predicted point as the center. The vertical axis is a number of times the position standard deviation, and the horizontal axis is a number of times the obstacle size. The predicted risk region is then transformed back to each robot's local coordinate system to obtain the obstacle's spatiotemporal trajectory prediction result.

5. The robot dynamic obstacle avoidance path planning method according to claim 1, characterized in that, In the preprocessed image, a feature point detection algorithm is used to identify and extract fixed static landmarks. Based on the UWB label coordinates, the relative position vector of the target robot with respect to the static landmarks is calculated to construct a robot position fingerprint, which is used to correct positioning errors and verify the stationary state. This includes: The image is converted into a distortion-free top view using a camera calibration matrix; In the top view, a feature point detection algorithm is used to identify and extract fixed static landmarks, including wall corners, column corners, and fixed equipment corners; The relative position vector of the target robot with respect to a static landmark at a distance that meets the requirements is calculated based on the coordinates of the UWB tag. This vector serves as the robot's position fingerprint. If the difference in position fingerprints at different times is within a set tolerance range and the time interval exceeds the preset minimum judgment interval, then the robot is determined to be in a stationary state.

6. The robot dynamic obstacle avoidance path planning method according to claim 1, characterized in that, The method involves combining the obstacle spatiotemporal trajectory prediction results with the robot's position fingerprint, multi-robot collaborative perception results, and the corresponding trust weights of each robot to dynamically determine the obstacle avoidance strategy and replan the path on the planned path, generating collaborative scheduling instructions, including: Based on the obstacle spatiotemporal trajectory prediction results, the robot position fingerprint, the multi-robot collaborative perception results, and the corresponding trust weights of each robot, a conflict risk assessment is performed on each mobile robot, including: calculating the spatiotemporal overlap between the planned path of each mobile robot and all obstacle prediction risk areas, the estimated collision time and collision probability, and quantifying and generating a conflict risk level. Layered decision-making is performed. When a static obstacle is detected and its continuous tracking exceeds a set number of frames using the spatiotemporal database, it is marked as a fixed obstacle, the global environment map is updated, and a detour strategy is triggered in the path planning layer. When the predicted risk area of ​​a dynamic obstacle overlaps with the robot's planned path in both time and space and the collision time is less than a safety threshold, the probability of path conflict is calculated. If the probability exceeds the first-level threshold, a cooperative deceleration command is sent to the relevant robot. If it exceeds the second-level threshold, a distributed path replanning request is triggered. When a temporary occlusion is detected that persists within a set number of tracking frames but has no fixed features, it is marked as a pause and yield area. The robot is controlled to wait at a safe distance or initiate a local detour to form a cooperative scheduling command.

7. The robot dynamic obstacle avoidance path planning method according to claim 1, characterized in that, Based on the cooperative scheduling instructions, and combined with the robot's current state and environmental characteristics retrieved from the spatiotemporal database, the system performs local obstacle avoidance control or global task reallocation, and generates a structured execution log, including: Based on the collaborative scheduling instructions, the robot's current battery level, load weight, and motion status data are retrieved from the dynamic environment spatiotemporal database. A distributed negotiation mechanism is used to dynamically redistribute tasks, prioritizing the transfer of tasks in high-risk areas to low-conflict mobile robots. Among them, low-conflict mobile robots are those whose current planned path or predicted trajectory does not significantly overlap with the predicted risk area of ​​dynamic obstacles, whose collision probability is lower than a preset safety threshold, and which are not in a paused yielding or emergency braking state, based on obstacle spatiotemporal trajectory prediction results and conflict risk assessment. For obstacle avoidance control of each mobile robot, differentiated strategies are implemented based on the type of obstacle. Create a structured execution log, which includes obstacle ID, type, obstacle spatiotemporal trajectory prediction results, a list of mobile robot IDs participating in the collaboration, the trust weight of each robot, scheduling decision type, spatiotemporal stamp, and confidence score.

8. The robot dynamic obstacle avoidance path planning method according to claim 7, characterized in that, The differentiated strategies include performing geometrical maneuvers around static obstacles; performing speed matching and lateral avoidance for dynamic obstacles; and performing a wait-detect-and-go strategy for temporary obstructions.

9. A robot dynamic obstacle avoidance path planning system, characterized in that, include: The acquisition unit is used to acquire images, UWB tag coordinates, timestamps and body motion state data collected by edge computing terminals mounted on multiple mobile robots, and to build a dynamic environment spatiotemporal database based on the acquired data; The processing unit is used to preprocess the image, identify and classify static obstacles, dynamic obstacles and temporary occlusion targets, output obstacle feature information, confidence score and motion vector, obtain multi-robot collaborative perception results, and use spectral similarity calculation and correlation coefficient verification methods to perform interference removal and validity verification on cross-terminal data in the dynamic environment spatiotemporal database to obtain the corresponding confidence weight of each robot. The prediction unit is used to filter dynamic obstacles close to the robot's planned path and predict their future positions based on the multi-robot collaborative perception results. It constructs the dynamic obstacle's motion trajectory by associating cross-robot, continuous multi-frame data, and calibrates it in a unified world coordinate system to generate a predicted risk area, thus forming an obstacle spatiotemporal trajectory prediction result. The obstacle spatiotemporal trajectory prediction result is associated with the timestamps in the spatiotemporal database and the UWB tag coordinates. The fingerprint construction unit is used to identify and extract fixed static landmarks in the preprocessed image using a feature point detection algorithm, calculate the relative position vector of the target robot relative to the static landmarks based on the UWB label coordinates, and construct the robot position fingerprint to correct positioning errors and verify the stationary state. The instruction generation unit is used to dynamically determine the obstacle avoidance strategy and replan the path based on the obstacle spatiotemporal trajectory prediction result, the robot position fingerprint, the multi-robot collaborative perception result, and the corresponding trust weight of each robot, and generate collaborative scheduling instructions. The execution unit is used to perform local obstacle avoidance behavior control or global task reallocation based on the cooperative scheduling instructions and the robot's current state and environmental characteristics retrieved from the spatiotemporal database, and to generate a structured execution log.

10. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 9.

Images