An AI-based robot obstacle recognition method and system

Abstract

This invention relates to an artificial intelligence-based robot obstacle recognition method and system, comprising: collecting the material of the obstacle, acquiring the physical data of the material, collecting real-time data in the scene and matching it with the material database, identifying the material type of the current object, obtaining the deformation confidence coefficient based on deformation characteristics, calculating the deformation offset, adjusting the deformation offset to obtain a comprehensive deformation offset, recognizing the obstacle, and then planning the robot path. Through the obtained comprehensive deformation offset, high-precision recognition and safe path planning of flexible and rigid obstacles are achieved. By suppressing airflow disturbances and separating the interaction of flexible bodies, combined with deformation offset correction, the actual movement of flexible bodies can be more accurately identified.

Description

Technical Field

[0001] This invention relates to the field of robot obstacle recognition, and more specifically to a robot obstacle recognition method and system based on artificial intelligence. Background Technology

[0002] Intelligent material handling robots, as core equipment in the field of industrial automation, integrate autonomous navigation, environmental perception, and intelligent decision-making systems, and are specifically designed for material handling tasks in complex factory environments. In actual operation scenarios, robots need to identify two types of obstacles in real time in dynamically changing industrial environments: static obstacles (such as fixed equipment, walls, and shelves) and dynamic obstacles (such as moving AGVs, operators, and suspended flexible workpieces).

[0003] Existing patented technologies all assume that obstacles are rigid bodies. However, flexible objects (such as hanging plastic curtains or liquid leaks) are common in industrial scenarios. Their deformation can cause nonlinear noise in optical flow vectors and lidar data, which traditional methods cannot model. The deformation of flexible bodies is easily misjudged as dynamic obstacles. Traditional rigid body assumption models cannot handle nonlinear deformations of fabrics, plastic curtains, etc. There is insufficient multimodal data fusion, lidar and vision sensors operate in isolation, and vibration superposition when multiple robots cooperate can easily cause resonance of flexible bodies. The lack of cluster control leads to conflict in task efficiency, and the lack of airflow disturbance compensation causes flexible bodies to undergo secondary deformation due to the influence of robot wake. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an artificial intelligence-based robot obstacle recognition method and system. By obtaining the comprehensive deformation offset, it achieves high-precision recognition and safe path planning for both flexible and rigid obstacles. Through airflow disturbance suppression and flexible body interaction separation, combined with deformation offset correction, it can more accurately identify the actual movement of the flexible body.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a robot obstacle recognition method based on artificial intelligence.

[0006] S101: Collect the material of obstacles, obtain physical data of the materials, establish a material database based on the physical data, collect real-time data in the scene and match it with the material database, identify the material type of the current object, extract the deformation characteristics of non-rigid objects based on the material type, and obtain the deformation confidence coefficient.

[0007] S102, Calculate the deformation offset based on physical data and deformation confidence coefficient;

[0008] S103 uses 3D point cloud data to adjust the deformation offset to obtain a comprehensive deformation offset. The comprehensive deformation offset is used to correct the optical flow vector. Based on the corrected optical flow vector and LiDAR data, obstacles are identified, and then the robot path is planned.

[0009] S104, acquire real-time data, establish an airflow model based on the real-time data, obtain the disturbance vector according to the airflow model, and set up an airflow compensation device to compensate for the disturbance vector;

[0010] S105, Collect multimodal data and vibration data, identify and segment the flexible body based on the multimodal data, extract the deformation features of the segmented flexible body, calculate the momentum transfer based on the deformation features, and then separate the actual motion of the flexible body. Correct the comprehensive deformation offset based on the actual motion of the flexible body. The multimodal data includes lidar data and image data.

[0011] Preferably, the physical data includes Young's modulus and viscosity coefficient. Young's modulus is a physical quantity describing the stiffness of a material, defined as the ratio of normal stress to normal strain during the elastic deformation stage of the material. Viscosity coefficient is a physical quantity describing the fluid's ability to resist shear flow, defined as the shear stress under a unit velocity gradient between fluid layers.

[0012] Preferably, the formula for calculating the deformation offset is: in, This is the deformation offset. The deformation confidence coefficient represents the probability that an object is flexible. Young's modulus is the property of a material to resist elastic deformation. Strain energy density represents the elastic potential energy stored per unit volume. The rate of change of strain energy density is the energy source driving elastic deformation. The viscosity coefficient characterizes a material's ability to resist viscous flow. Let be the velocity gradient tensor, and let be the Jacobian matrix of the velocity field on the object's surface.

[0013] Preferably, the formula for obtaining the comprehensive deformation offset based on the 3D point cloud data is as follows: ,in, The overall deformation offset is used for correction, where, The overall deformation offset is used to correct motion errors in optical flow and lidar data. This is the deformation offset. This is the deformation confidence coefficient. For point cloud data, The rate of change of point cloud density represents the amount of change in point cloud density per unit time. As a baseline point cloud density, , and All of these are weighting coefficients, determined based on experiments.

[0014] Preferably, obstacles are identified using a joint probability model, which includes motion probability, geometric probability, and a soft suppression term. A threshold range is set based on the three probabilities: motion probability, geometric probability, and soft suppression term. The threshold range includes a maximum threshold and a minimum threshold. If the total probability exceeds the maximum threshold, it is identified as an obstacle; if it is below the minimum threshold, it is identified as background; if it is between the minimum and maximum thresholds, it is further confirmed using a semantic segmentation model.

[0015] Preferably,

[0016] S201 collects vibration data in real time and identifies vibration sources generated by different robots based on the vibration data;

[0017] S202, Collect robot status data, establish a shared communication mechanism through the status data of multiple robots, and plan the paths of multiple robots according to the shared communication mechanism.

[0018] Preferably, the formula for calculating the radius of influence based on the vibration source is: ,in, The radius of influence represents the distance from the vibration source to the flexible body. For material density, The vibration damping coefficient describes the rate at which the vibration amplitude decays exponentially with time. For Young's modulus, The initial amplitude was obtained by extracting vibration characteristics. The threshold amplitude for flexible body deformation is the vibration amplitude corresponding to the maximum safe deformation allowed by the flexible body.

[0019] Preferably, a time-sharing start strategy is used to plan the paths of multiple robots, wherein the time-sharing start strategy is to plan the paths by controlling the start time interval and sequence of multiple robots.

[0020] Preferably,

[0021] S301 generates groups based on collected vibration data and task requirements, calculates vibration suppression intensity based on the robots within the group, and formulates vibration suppression strategies within the group based on the vibration suppression intensity.

[0022] S302 manages intergroup buffer zones by monitoring cross-group vibration.

[0023] An artificial intelligence-based robot obstacle recognition system, the system comprising: a data module, a deformation module, a path planning module, and a grouping module;

[0024] The data module is used to collect robot status data and vibration data;

[0025] The deformation module is used to collect physical data of obstacle materials, establish a material database, extract deformation features, and calculate deformation confidence.

[0026] The path planning module is used to identify obstacles and plan paths according to path algorithms;

[0027] The grouping module generates groups based on physical distance, task coupling degree, and vibration homogeneity, selects the master node robot, calculates the vibration suppression intensity within the group, and adjusts the speed of the group members.

[0028] The beneficial effects of this invention are: by obtaining the comprehensive deformation offset, high-precision identification and safe path planning of flexible and rigid obstacles can be achieved; by suppressing airflow disturbance and separating the interaction of flexible bodies, combined with deformation offset correction, the actual movement of flexible bodies can be identified more accurately, improving the accuracy of obstacle identification of intelligent handling robots in complex environments (with airflow disturbance and multiple flexible body interaction scenarios), ensuring that the robot can accurately identify and avoid obstacles, and guaranteeing the stability and safety of robot operation;

[0029] By introducing vibration source differentiation and cluster collaborative control, combined with multi-sensor fusion and machine learning technology, the system achieves accurate identification of flexible body deformation and active suppression of vibration transmission in a multi-robot cluster environment. Through vibration source differentiation and cluster collaborative control, it reduces misjudgment of flexible body deformation caused by vibration transmission, improves obstacle recognition accuracy, optimizes the motion coordination of multiple robots, reduces the risk of flexible body deformation, and enhances the stability of cluster operation.

[0030] By dynamically grouping and collaboratively suppressing vibrations, multiple robots can operate with low vibration and high efficiency in complex industrial scenarios, while ensuring the safety of flexible bodies. Through vibration homogeneous grouping and intra-group vibration suppression strategies, the system upgrades from passive vibration avoidance to active neutralization. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating an artificial intelligence-based robot obstacle recognition method according to the present invention.

[0032] Figure 2 This is a schematic diagram illustrating the process of identifying vibration sources generated by different robots based on vibration data according to the present invention.

[0033] Figure 3 This is a schematic diagram of the process for generating groups according to the present invention;

[0034] Figure 4 This is a block diagram of an artificial intelligence-based robot obstacle recognition system according to the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0037] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0038] Example 1: Figure 1 This is a flowchart illustrating an artificial intelligence-based robot obstacle recognition method according to Embodiment 1 of the present invention, comprising the following steps:

[0039] S101: Collect the material of obstacles, obtain physical data of the materials, establish a material database based on the physical data, collect real-time data in the scene and match it with the material database, identify the material type of the current object, extract the deformation characteristics of non-rigid objects based on the material type, and obtain the deformation confidence coefficient.

[0040] Furthermore, the materials of flexible, viscous, and rigid objects in the robot's operating environment are collected. The acquired physical data includes Young's modulus and viscosity coefficient. Young's modulus is a physical quantity describing the stiffness of a material, defined as the ratio of normal stress to normal strain during the elastic deformation stage. Viscosity coefficient is a physical quantity describing the fluid's resistance to shear flow, defined as the shear stress under a unit velocity gradient between fluid layers. Pressure is applied to the collected materials through tensile tests, and stress-strain curves are measured to calculate the elastic modulus, i.e., Young's modulus. The flow resistance of the collected materials is measured using a viscometer, and the viscosity coefficient is calculated from the flow resistance. The obtained physical data is entered into a SQL database to establish a material database. The database fields include material name, Young's modulus, viscosity coefficient, and classification label. The materials are classified according to deformation type into flexible (fabric), viscous (plastic curtain), and rigid (metal).

[0041] A 32-line Ouster OS1-3 LiDAR was installed at the center of the robot's front end. The LiDAR's scanning frequency was set to 10-20Hz. The LiDAR scanned the surfaces of objects in the scene in real time, generating 3D point cloud data containing the spatial coordinates and reflection intensity of each point. The 3D point cloud data was preprocessed. For each point, its K nearest neighbors were selected, a local surface was fitted, and the principal curvature was calculated. A quadratic surface was fitted using the least squares method, and the principal curvature was calculated using the eigenvalues ​​of the Hessian matrix of the surface equation. Reflection intensity was weighted, and the curvature distribution characteristics were obtained by calculating the mean and variance of all points in the point cloud. A curvature histogram was generated based on the curvature distribution characteristics, and a Basler global shutter industrial camera was used. The acA1920-50gc camera acquires continuous frame images of the object's surface in real time. It is installed at the same angle as the LiDAR, and the LiDAR and camera are triggered synchronously. The Lucas-Kanade algorithm is used to calculate the optical flow field of two consecutive frames. The optical flow field is arranged according to time to form an optical flow time series. The optical flow time series is converted from a time-domain signal to a frequency-domain signal using Fourier transform. Rigid objects (low-frequency vibration) and flexible objects (high-frequency vibration) are distinguished based on the object's vibration frequency characteristics. The curvature distribution characteristics and vibration frequency characteristics are concatenated into a comprehensive feature. The KNN (K-Nearest Neighbors) algorithm is used to calculate the similarity between the comprehensive feature and the materials in the material database. A similarity threshold is set based on historical data. If the calculated similarity is greater than the threshold, the real-time data is considered to have successfully matched the material in the material database; otherwise, the match is unsuccessful and the material is marked as unknown, thus identifying the material type of the current substance.

[0042] Based on the identified material type, deformation features are extracted to determine the object's flexibility. A deformation filter based on a convolutional neural network (CNN) is then set, comprising an input layer, convolutional layers, pooling layers, and a fully connected layer. A physics engine, Blender, is used to synthesize fabric swaying animations and generate simulated data. The deformation filter is trained using this simulated data. A Basler acA1920-50gc industrial camera captures continuous frame images of the object's surface in real time, converting the RGB images to single-channel grayscale values. The grayscale calculation formula is existing technology and will not be elaborated here. The grayscale image is input to the deformation filter's input layer. The convolutional layer extracts low-level edge features using convolutional kernels based on the input grayscale image and combines these low-level features into high-level oscillation patterns. The pooling layer reduces the feature map size through max pooling. The fully connected layer expands the feature map into a vector and outputs the deformation confidence coefficient C. d If C d =0, the object is rigid and has no edge oscillations; if C d =1, the object is flexible, and the edge oscillation is significant.

[0043] S102, Calculate the deformation offset based on physical data and deformation confidence coefficient;

[0044] Specifically, based on the real-time material classification results, the physical data of the corresponding material is obtained from the material database. Based on the Young's modulus and viscosity coefficient in the physical data (Young's modulus represents the material's resistance to elastic deformation, and viscosity coefficient represents the material's resistance to viscous flow), the deformation offset is calculated using the obtained Young's modulus and viscosity coefficient, using the following formula: ,in, This is the deformation offset, which quantifies the current degree of deformation of the object. The larger the value, the more significant the deformation. The deformation confidence coefficient represents the probability that an object is flexible. Young's modulus is a property that characterizes a material's resistance to elastic deformation; the higher the value, the more rigid the material. Strain energy density represents the elastic potential energy stored per unit volume. The rate of change of strain energy density is the energy source driving elastic deformation. The viscosity coefficient characterizes a material's ability to resist viscous flow; the higher the value, the more viscous the material. Let be the velocity gradient tensor, and be the Jacobian matrix of the velocity field on the object's surface. The deformation offset is calculated using the above formula, and the deformation offset represents the degree of significance of the material deformation.

[0045] S103, the deformation offset is adjusted according to the three-dimensional point cloud data to obtain the comprehensive deformation offset. The comprehensive deformation offset is used to correct the optical flow vector. Based on the corrected optical flow vector and the data of the lidar, obstacles are identified and the robot path is planned.

[0046] Furthermore, based on the 3D point cloud data of the scene collected by the LiDAR, the deformation offset is adjusted, and the adjusted formula is: ,in, The overall deformation offset is used for correction, where, The overall deformation offset is used to correct motion errors in optical flow and lidar data. This is the deformation offset. This is the deformation confidence coefficient. For point cloud data, The rate of change of point cloud density represents the amount of change in point cloud density per unit time, reflecting the degree of deformation of an object's surface. As a baseline point cloud density, , and All of these are weighting coefficients, determined based on experiments.

[0047] The Farneback algorithm is used to calculate the full-pixel optical flow field of adjacent grayscale images to obtain the original optical flow vector. Deformation of flexible objects can introduce spurious motion components into the optical flow vector. By quantifying the degree of local deformation through comprehensive deformation offset, the optical flow field is weighted and suppressed, preserving the true motion components. The formula for correcting the optical flow vector is as follows: ,in, The corrected optical flow vector field represents the true motion vector at pixel coordinates (x, y) after deformation interference suppression. The original optical flow vector field is the uncorrected motion vector directly calculated by the optical flow algorithm, containing a mixture of real motion and deformation disturbance components. The comprehensive deformation offset is a normalized index representing the degree of deformation of a flexible object at the quantized pixel coordinates (x, y). The optical flow amplitude is obtained based on the corrected optical flow vector field. If the optical flow amplitude is greater than the set threshold, it is marked as a candidate motion vector. For each candidate motion vector, the distance of the corresponding LiDAR point is queried. If the distance of the LiDAR point changes more than the distance change threshold within a certain period of time, it is judged as a moving object vector; otherwise, it is a stationary object vector.

[0048] Based on each vector in the corrected optical flow vector, its motion features, such as velocity and direction of motion, are extracted. Using the distance information and reflection intensity of the object extracted by the lidar, the optical flow motion features and lidar geometric information are input into a joint probability model to comprehensively calculate the probability that each region is an obstacle. Finally, an obstacle label is output. The joint probability model includes motion probability, geometric probability, and a flexibility suppression term. The motion probability calculates the likelihood that an object is a dynamic obstacle based on its speed; the faster the speed and the closer it is to a preset threshold, the higher the probability. The geometric probability combines lidar distance data; if the object is within a safe distance, then... If an object is identified as a high-probability obstacle, its probability is reduced if it is far away or if there is no detection data. The flexibility suppression term targets flexible objects (such as curtains and flags). If the deformation confidence is high and the deformation offset is large, the probability of it being identified as an obstacle is significantly reduced, avoiding false alarms. A threshold range is set by combining the probabilities of motion, geometry, and flexibility suppression. The threshold range includes a maximum threshold and a minimum threshold. If the total probability exceeds the maximum threshold, it is confirmed as an obstacle; if it is below the minimum threshold, it is identified as background; if it is between the minimum and maximum thresholds, it is further confirmed by a semantic segmentation model. In the known map, A is used. The algorithm plans the shortest path to ensure global optimality. When an obstacle enters the safe zone, it switches to RRT (Reverse Response Time). The algorithm quickly generates feasible paths within a local range, adapts to dynamic changes, and combines A Global efficiency and RRT The robot exhibits dynamic adaptability, prioritizing the global path while flexibly avoiding obstacles locally. It uses cubic Bézier curves to fit the optimized path, eliminating sharp turns and ensuring smooth robot movement.

[0049] S104, acquire real-time data, establish an airflow model based on the real-time data, obtain the disturbance vector according to the airflow model, and set up an airflow compensation device to compensate for the disturbance vector;

[0050] Specifically, the acquired real-time data includes robot speed, shape, and surrounding environment data. A speed sensor is used to acquire the robot's motion speed in real time, a 3D scanning device is used to acquire the robot's shape data, and environmental sensors are used to collect information about the surrounding environment, including temperature, humidity, and air density. Based on the principles of computational fluid dynamics (CFD), a Realizable k-ε airflow model is built using FLUENT software. The acquired robot speed, shape, and surrounding environment data are input into the airflow model. Based on the input data and internal algorithm calculations, ε outputs the predicted distribution of airflow around the robot during its movement, including the values ​​of parameters such as airflow velocity, direction, and pressure at different locations, i.e., the disturbance vector. Multiple tests are conducted in a simulated environment, comparing the simulation results with actual measurement data until the model's prediction accuracy meets the requirements. Based on the size and layout of the robot, LiDAR, and image acquisition equipment, an airflow guide plate is selected as the airflow compensation device. The airflow direction and disturbance area are determined based on the output of the airflow model. The airflow guide plate is installed in the disturbance area, and an initial angle range is set. The angle selection is based on the airflow direction and the degree of change required. Displacement sensors, acceleration sensors, and angle sensors are installed around the suspended flexible body, collecting sensor data at a frequency of 10 times per second. The collected data includes real-time values ​​of displacement, acceleration, and angle, and the collection time is recorded. The motion state parameters of the flexible body are calculated using a data analysis algorithm. The real-time collected data is compared with preset normal motion state parameters. When the swing amplitude of the flexible body exceeds a preset threshold, it is determined that the airflow disturbance has a significant impact. The angle of the airflow guide plate is adjusted to compensate for the influence of the airflow generated by the robot's movement on the suspended flexible body.

[0051] S105: Collect multimodal data, identify and segment the flexible body based on the multimodal data, extract the deformation features of the segmented flexible body, calculate the momentum transfer based on the deformation features, and then separate the real motion of the flexible body. Correct the comprehensive deformation offset based on the real motion of the flexible body.

[0052] Furthermore, the collected multimodal data includes LiDAR data and image data. For LiDAR data acquisition, point cloud data of the scene is obtained by setting the scanning range and scanning frequency of the LiDAR. The point cloud data contains the three-dimensional coordinate information of objects. High-definition cameras are used to capture scene images, obtaining visual information such as the color and texture of objects. The acquired images are processed using the Mask R-CNN model, and the labeled flexible body image data is used to apply the Mask R-CNN model. The model is trained, during which it learns the characteristics of different types of flexible bodies (such as fabric and dust curtains), including color, texture, and shape. By continuously adjusting the model's parameters, the accuracy of the model in recognizing flexible bodies is improved. The trained model is then applied to the acquired images to identify flexible body regions in the images and classify different types of flexible bodies. The Euclidean clustering algorithm is used to analyze the LiDAR point cloud data, clustering point clouds belonging to the same object together. Combined with the image recognition results, the point cloud corresponding to the flexible body is segmented from the scene point cloud. For the segmented flexible body, deformation feature parameters are extracted, and the optical flow algorithm is used to calculate the motion vector of each point on the surface of the flexible body to obtain the deformation amplitude and deformation frequency.

[0053] The mass and elastic modulus of the flexible body were obtained through experimental measurements. A momentum transfer calculation model, the Kelvin-Voigt model, was established based on Newton's second law and the law of conservation of momentum. The model incorporates the deformation characteristic parameters, mass, and elastic modulus of the flexible body to calculate the momentum transfer between the various flexible bodies. The Kelvin-Voigt model outputs the momentum transfer calculation results. Based on these results, kinematic equations were established. These equations describe the relationship between the motion state of the flexible body and momentum transfer. By establishing these equations, the spurious motion caused by momentum transfer can be separated from the total motion. Solving these equations allows for the identification of further spurious motions. The kinetic equations are used to obtain the actual motion of each flexible body caused only by its own deformation. Based on the actual motion of each flexible body, the comprehensive deformation offset is corrected. The cumulative effect of instantaneous velocity in momentum transfer over time affects the comprehensive deformation offset. The instantaneous velocity is obtained by using momentum conservation as the result of momentum transfer. The instantaneous velocity is then integrated over time to obtain the offset caused by momentum. The comprehensive deformation offset is corrected based on the offset caused by momentum. The corrected deformation offset = comprehensive deformation offset - offset caused by momentum. The corrected deformation offset is then substituted into step S103 to correct the optical flow vector, further improving the efficiency of obstacle recognition and path planning.

[0054] The technical solutions in the above embodiments of this application have at least the following technical effects or advantages: by obtaining the comprehensive deformation offset, high-precision identification and safe path planning of flexible and rigid obstacles can be achieved; by suppressing airflow disturbances and separating flexible body interactions, combined with deformation offset correction, the real motion of flexible bodies can be identified more accurately, improving the accuracy of obstacle identification of intelligent handling robots in complex environments (where there are airflow disturbances and multiple flexible body interaction scenarios), ensuring that the robot can accurately identify and avoid obstacles, and guaranteeing the stability and safety of robot operation.

[0055] Example 2: Based on the description of a single robot in Example 1, this example improves the accuracy of obstacle recognition by identifying vibration sources generated by multiple robots during operation. Figure 2 As shown.

[0056] S201 collects vibration data in real time and identifies vibration sources generated by different robots based on the vibration data;

[0057] Furthermore, three vibration sensors are installed on the robot, fixed along the X / Y / Z axes of the robot's frame (e.g., at the chassis, drive wheel brackets, and load platform connections). This allows the sensors to collect three-dimensional vibration signals. Based on the vibration data collected by the installed sensors, the collected vibration data undergoes noise reduction processing. 5G communication enables time synchronization of all sensors. Bandpass filtering is applied to the collected vibration signals using a fourth-order Butterworth filter with a set cutoff frequency. A bidirectional filtering method is employed, using zero-phase filtering technology, first forward filtering and then reverse filtering, to ensure the signal's time-domain waveform remains undistorted. The filter cutoff frequency is dynamically adjusted based on real-time spectrum analysis according to changes in the robot's operating conditions. Vibration features are extracted from each feature window, with T=0.1... s is the window length, which corresponds to 100 sampling points. Adjacent windows overlap by 50 points to avoid missing abrupt changes in features. If the signal ends with less than one window length, zeros are padded to make it an integer multiple of the window length. A Fourier transform is performed on the windowed time-domain signal to obtain a complex spectrum. A single-sided spectrum (0~200 Hz) is taken, retaining only the positive frequency portion. The amplitude is multiplied by 2 to compensate for energy loss. The frequency point with the largest amplitude in the range of 10~200 Hz is found as the dominant frequency of the current window. Based on the spectral amplitude corresponding to the dominant frequency, combined with the sensor sensitivity and range, it is converted into an actual physical quantity, i.e., amplitude. The phase angle of the complex spectrum corresponding to the dominant frequency is taken to reflect the phase difference between the vibration signal and the reference signal. The extracted dominant frequency, amplitude, and phase are normalized to eliminate dimensional differences.

[0058] Vibration sources are identified using a machine learning model called Support Vector Machine (SVM). Historical datasets are collected, including vibration characteristics of different robots under various motion modes. The SVM is trained based on these historical datasets. Real-time characteristic vectors such as the dominant frequency, amplitude, and phase are input into the SVM. The SVM outputs classification results, assigning one vibration source to one robot. The radius of influence is then calculated based on the vibration source using the following formula: ,in, The radius of influence represents the distance from the vibration source to the flexible body. For material density, The vibration attenuation coefficient describes the rate at which the vibration amplitude decays exponentially with time or distance. For Young's modulus, The initial amplitude was obtained by extracting vibration characteristics. The vibration amplitude is the threshold amplitude for the deformation of the flexible body, which is the vibration amplitude corresponding to the maximum safe deformation allowed by the flexible body. Based on the calculated radius of influence, the range of influence of the vibration on the flexible body is determined.

[0059] S202, Collect robot status data, establish a shared communication mechanism through the status data of multiple robots, and plan the paths of multiple robots according to the shared communication mechanism;

[0060] Specifically, the robot's state data, including position, velocity, and vibration characteristics, is collected. This state data is shared using 5G URLLC (Ultra-Reliable Low-Latency Communication) mode, establishing a shared communication mechanism. Based on the shared information, a time-sharing startup strategy is used to plan the path. This strategy controls the startup time interval and sequence of multiple robots to avoid vibration superposition causing resonance or excessive deformation of the flexible body, thus ensuring the safety of path planning. The vibration amplitude at time t is calculated based on the initial vibration amplitude using the following formula: ,in, Let A0 be the vibration amplitude at time t, and A0 be the initial vibration amplitude. Here, A(t) is the damping coefficient, and t is the time variable. To avoid the superposition of vibrations from adjacent robots, it is necessary to ensure that the vibration amplitude of the previous robot decays to less than 10% of its initial value, i.e., A(t) ≤ 0.1A0. Substituting this into the equation... From: Taking the logarithm of both sides of the above formula and simplifying, we get... That is, adjacent robots need to start at least 0.5 seconds apart. They are sorted from highest to lowest according to the urgency of the task. When the priorities are the same, the robot that is farther away from the flexible body starts first. If a robot is delayed in starting because it is waiting for the vibration of the previous robot to decay, the central controller replans its path, shortens the actual travel distance, and compensates for the time loss. In the path planning, the vibration influence range of the already started robot is marked as a dynamic obstacle, and subsequent robots need to avoid these areas.

[0061] The technical solutions in the above embodiments of this application have at least the following technical effects or advantages: by introducing vibration source differentiation and cluster collaborative control, combined with multi-sensor fusion and machine learning technology, accurate identification of flexible body deformation and active suppression of vibration transmission are achieved in the environment of multiple robot clusters. Through vibration source differentiation and cluster collaborative control, the misjudgment of flexible body deformation caused by vibration transmission is reduced, the accuracy of obstacle recognition is improved, the coordination of multiple robot movements is optimized, the risk of flexible body deformation is reduced, and the stability of cluster operation is enhanced.

[0062] Example 3: Based on Examples 1 and 2 above, this example uses dynamic grouping and cooperative suppression to reduce vibration interference while ensuring task efficiency and flexible body safety, such as... Figure 3 As shown.

[0063] S301 generates groups based on collected vibration data and task requirements, calculates vibration suppression intensity based on the robots within the group, and formulates vibration suppression strategies within the group based on the vibration suppression intensity.

[0064] Specifically, the grouping involves combining multiple robots into a highly efficient, collaborative unit, while avoiding efficiency degradation due to vibration interference or task conflicts. First, filtering is performed based on physical distance. The central control system acquires the current position and task requirements of all robots in real time. A circular area with a radius of 10 meters is defined centered on the order task's starting point. The number of robots within this area is counted, and the straight-line distance between any two robots is calculated. If the number of robots in the area is ≥2, and the distance between any two robots is ≤10 meters, they are grouped into a physical proximity candidate group. If the number of robots in the area is insufficient, the search radius is expanded to 15 meters, but marking is required. This is denoted as a low-density candidate group, and several physical proximity candidate groups are generated. The remaining execution time of each robot's current task is obtained from the system. For each physical proximity candidate group, the remaining task time difference between all robots in the group is calculated, i.e., the difference in remaining task time between any two robots. If the remaining task time difference between two robots in the group is greater than 5 seconds, their task coupling is considered insufficient, and the robot with the longest remaining task time is removed. If the number of robots in the group is less than 2 after removal, the group is disbanded, and the robot is marked as a robot to be merged. Task coupling candidate groups are generated based on the remaining task time difference, with the remaining task time difference of robots in each group ≤ 5 seconds to ensure motion synchronization. Historical vibration data is acquired from the robot's vibration sensors, and the dominant vibration frequency of each robot in the group is statistically analyzed. The average dominant vibration frequency of all robots in the group is calculated, and the frequency difference between any two robots is calculated. If the frequency difference between two robots is greater than 3 Hz, their vibration homology is considered insufficient. If the vibration frequency difference within a group is large, the group is split into two subgroups. If only a few robots deviate from the dominant frequency, only that robot is removed, and the remaining members are retained. The vibration homology is re-tested for the split subgroups or new groups until all groups meet the conditions. The final grouping is generated based on the above physical distance screening, task coupling degree, and vibration homology screening. Robots in each group are physically adjacent to each other. The tasks are coupled and the vibrations originate from the same source, providing a basic condition for efficient collaboration. In the generated group, the master node is identified by the vibration transmission coefficient. The master node is used to coordinate actions within the group, calculate vibration damping parameters, and handle emergencies. The vibration transmission coefficient is calculated by statistically analyzing the average vibration attenuation contribution of the robot to other robots over a period of time. The vibration attenuation contribution is calculated by the proportion of vibration amplitude reduced by the active vibration reduction algorithm. The transmission coefficients of all robots in the group are sorted, and the one with the highest coefficient is elected as the master node. The central control system sends the grouping rules and master node permissions to each robot. The grouping rules include thresholds for physical proximity, task coupling, and vibration homogeneity.

[0065] The master node robot collects the vibration attenuation rate of each member in real time, calculates the percentage decrease in vibration energy using sensor data, and then obtains the vibration attenuation rate of each member. Based on this vibration attenuation rate, the vibration suppression intensity is calculated using the following formula: ,in, Indicates vibration damping strength. This represents the vibration attenuation rate of the i-th group member. Indicates the number of group members. The master node authority coefficient is a pre-set coefficient used to adjust the influence of the master node in vibration suppression calculation. The speed of the group members is adjusted according to the calculated vibration suppression intensity. The vibration suppression intensity is multiplied by the reference speed to obtain the actual speed of the group members. The reference speed is a pre-set running speed of the group members under normal conditions, which is a fixed reference value. The vibration suppression intensity is compared with the threshold. If the vibration suppression intensity is less than the threshold, the master node is triggered to recalculate or expand the grouping range.

[0066] S302 manages inter-group buffer zones by monitoring cross-group vibration;

[0067] Furthermore, the buffer is a specific memory area, essentially a container object for readable and writable data. The system obtains the conductance by monitoring the vibration of different groups. The conductance is calculated by multiplying the source group vibration energy, the target group receiving coefficient, and the buffer attenuation factor. The source group vibration energy refers to the vibration energy of the group of robots that generates vibration and conducts it outward. This energy value is measured and calculated by corresponding sensors. The target group receiving coefficient is a pre-set parameter that reflects the target group robots' ability to receive the source group vibration energy. Different target groups have different receiving coefficients due to their own structure, materials, and other factors. The buffer attenuation factor depends on the buffer settings, including the buffer's material, width, shape, and other factors, and is used to measure the buffer's ability to receive vibration energy. The system utilizes the attenuation effect of kinetic energy. A conductivity threshold is set, and the obtained conductivity is compared to this threshold. If the conductivity is greater than the threshold, it indicates that the amount of vibration energy from the source group transmitted to the target group through the buffer is relatively large, which will have an adverse effect on the target group. The system monitors the grouping spacing in real time. When the grouping spacing exceeds the set threshold, the edge robot, a special robot individual located at the robot grouping boundary, acts as a buffer node for inter-group interaction. It continuously monitors changes in vibration amplitude. If the attenuation is insufficient, the buffer width is expanded. The buffer width is dynamically adjusted according to changes in vibration amplitude. When the vibration sensor detects that the conductivity exceeds the threshold, the group boundary is adjusted. Based on conductivity interference, the physical proximity, task coupling, and vibration homology rules are re-matched for regrouping.

[0068] The technical solutions in the above embodiments of this application have at least the following technical effects or advantages: by dynamic grouping and collaborative vibration suppression, multiple robots can operate with low vibration and high efficiency in complex industrial scenarios, while ensuring the safety of flexible bodies. The upgrade from passive vibration avoidance to active neutralization is achieved through vibration homogeneous grouping and intra-group vibration suppression strategies.

[0069] Example 4: Figure 4 As shown, this embodiment provides an artificial intelligence-based robot obstacle recognition system, including: a data module, a deformation module, a path planning module, and a grouping module. The data module is used to collect robot state data and vibration data. The deformation module is used to collect physical data of obstacle materials, establish a material database, extract deformation features, and calculate deformation confidence. The path planning module is used to identify obstacles and plan paths according to path algorithms. The grouping module generates groups based on physical distance, task coupling degree, and vibration homogeneity, selects a master node robot, calculates vibration suppression intensity within the group, and adjusts the speed of group members. The data module is connected to the deformation module, the deformation module is connected to the path planning module, and the path planning module is connected to the grouping module.

[0070] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0071] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0072] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0073] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0074] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0075] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0076] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An artificial intelligence-based robot obstacle recognition method, characterized in that, S101, collect the material of the obstacle, obtain the physical data of the material, establish a material database based on the physical data, match the real-time data in the scene with the material database, identify the material type of the current object, extract the deformation feature of the non-rigid object based on the material type, and obtain the deformation confidence coefficient; S102, calculate the deformation offset according to the physical data and the deformation confidence coefficient; S103, adjust the deformation offset using three-dimensional point cloud data to obtain a comprehensive deformation offset, correct the optical flow vector using the comprehensive deformation offset, and identify the obstacle based on the corrected optical flow vector and the data of the laser radar, and further plan the robot path; S104, obtain real-time data, establish an air flow model based on the real-time data, obtain a disturbance vector according to the air flow model, and set an air flow compensation device to compensate for the disturbance vector; S105, collect multi-modal data and vibration data, identify and segment the flexible body according to the multi-modal data, extract the deformation feature of the segmented flexible body, calculate the momentum transfer according to the deformation feature, and separate the real motion of the flexible body, and correct the comprehensive deformation offset according to the real motion of the flexible body, wherein the multi-modal data includes laser radar data and image data.

2. The method of claim 1, wherein the method is based on artificial intelligence. The physical data includes Young's modulus and viscosity coefficient, the Young's modulus is a physical quantity describing the stiffness of the material, defined as the ratio of normal stress to normal strain in the elastic deformation stage; the viscosity coefficient is a physical quantity describing the ability of fluid to resist shear flow, defined as the shear stress under the unit velocity gradient between fluid layers. 3.The method of claim 1, wherein, The formula for calculating the deformation offset is: wherein, is the deformation offset, is a deformation confidence coefficient, representing the probability that the object is flexible, is the Young's modulus, representing the ability of the material to resist elastic deformation, is the strain energy density, representing the elastic potential energy stored in unit volume, is the rate of change of strain energy density, the energy source driving elastic deformation, is the viscous coefficient, representing the ability of the material to resist viscous flow, is the velocity gradient tensor, and is the Jacobian matrix of the object surface velocity field.

4. The method of claim 1, wherein the method is based on artificial intelligence. The formula for obtaining the comprehensive deformation offset from the three-dimensional point cloud data is: wherein, is the comprehensive deformation offset, used for correcting the motion error in the optical flow and the lidar data, is the deformation offset, is the deformation confidence coefficient, is the point cloud data, is the point cloud density change rate, indicating the change amount of the point cloud density per unit time, is the reference point cloud density, , and are all weight coefficients, determined according to experiments.

5. The method of claim 1, wherein, The obstacle is identified by a joint probability model, the joint probability model includes motion probability, geometric probability and flexible suppression term, a threshold range is set according to the three probabilities of motion probability, geometric probability and flexible suppression term, the threshold range includes a highest threshold and a lowest threshold, if the total probability exceeds the highest threshold, it is confirmed as an obstacle; If it is lower than the lowest threshold, it is determined as background, if it is between the lowest threshold and the highest threshold, it is further confirmed by a semantic segmentation model.

6. The artificial intelligence-based robot obstacle recognition method according to claim 1, characterized in that, S201, real-time collection of vibration data, identification of vibration sources generated by different robots according to the vibration data; S202, collect robot state data, multiple robots establish a shared communication mechanism through the state data, and plan the paths of multiple robots according to the shared communication mechanism.

7. The method of claim 6, wherein the method further comprises: The formula for calculating the influence radius of the vibration source is: wherein, is the influence radius, represents the distance from the vibration source to the flexible body, is the material density, is the vibration attenuation coefficient, describes the rate of exponential decay of the vibration amplitude over time, is the Young's modulus, is the initial amplitude, obtained by extracting the vibration features, is the flexible body deformation threshold amplitude, the maximum safe deformation of the flexible body corresponds to the vibration amplitude.

8. The method of claim 6, wherein the method is based on artificial intelligence. The paths of multiple robots are planned by a time-sharing start strategy, the time-sharing start strategy is to plan the paths by controlling the start time interval and sequence of multiple robots.

9. The artificial intelligence-based robot obstacle recognition method according to claim 6, characterized in that, S301, generating a grouping according to the collected vibration data and task requirements, calculating the vibration suppression strength based on the robots in the grouping, and formulating a vibration suppression strategy within the group according to the vibration suppression strength; S302, manage the inter-group buffer by monitoring the cross-group vibration.

10. An artificial intelligence-based robot obstacle recognition system applied to an artificial intelligence-based robot obstacle recognition method according to any one of claims 6-9, characterized in that, The system comprises a data module, a deformation module, a path planning module and a grouping module; The data module is used for collecting robot state data and vibration data; The deformation module is used for collecting obstacle material physical data, establishing a material database, extracting deformation characteristics and calculating deformation confidence; The path planning module is used for identifying obstacles and planning paths according to path algorithms; The grouping module generates a group according to physical distance, task coupling degree and vibration homology, selects a master robot, calculates the vibration suppression strength in the group and adjusts the speed of group members.

Images