Artificial intelligence based industrial robot control system and method

Abstract

This application discloses an artificial intelligence-based industrial robot control system and method. A digital twin elastic motion model of the target robot is constructed, and its equivalent elastic parameter distribution is dynamically updated; the spatiotemporal distribution characteristics of elastic deformation during multi-mode switching are predicted. Force mutation signals at the instant of each mode switching are estimated based on a generalized momentum observer. The non-parametrically modeled disturbance torque components are separated by a residual reconstructor, and the disturbance propagation path topology is calculated. The spatiotemporal distribution of elastic deformation and the disturbance propagation topology are converted into tensors of the same dimension and multiplied element-wise to generate a dynamic risk field. This risk field is modeled as a Gaussian random field, and a collision heatmap is obtained by sampling. This heatmap is input into a counterfactual reasoning model based on causal inference to generate a multi-level backward planning avoidance strategy tree. The optimal branch satisfying the operational constraints is selected, and a graded collision avoidance action is executed through an offline pre-trained risk-level response controller. This application can realize feedforward active collision avoidance control of elastic robots in highly dynamic and uncertain environments.

Description

Technical Field

[0001] This application relates to the field of industrial robot control technology, and more specifically, to an artificial intelligence-based industrial robot control system and method. Background Technology

[0002] Industrial robots with multiple switchable drive modes can switch between joint-driven, link-driven, and end effector-driven modes according to operational needs, exhibiting higher motion flexibility and load adaptability. They are widely used in complex industrial scenarios such as automobile manufacturing, electronic assembly, and logistics sorting. However, the kinematic and dynamic characteristics of the robot can change abruptly during multi-mode switching. At the same time, the robot body can experience unpredictable spatial drift due to elastic deformation and force changes during mode switching, leading to increased trajectory tracking errors. Furthermore, when robots work in a swarm, the motion trajectories of different robots overlap, which can easily cause collisions.

[0003] Traditional collision avoidance methods for industrial robots mainly employ static interferometry or passive collision avoidance based on preset rules. These methods cannot capture the dynamic motion changes of robot swarms in real time and ignore the high collision risk at mode-switching nodes, making them ill-suited for complex and ever-changing operational scenarios. Existing AI-based collision avoidance methods are mostly designed for robots with a single drive mode. Some methods rely on the statistical distribution of historical execution residuals for risk modeling, failing to effectively handle first-time or non-repetitive disturbances. Furthermore, existing methods often employ locally greedy collision avoidance strategies, failing to generate global feedforward plans that consider temporal causal relationships, easily leading to conflicts or secondary collisions between avoidance actions. Therefore, how to achieve feedforward active collision avoidance control for resilient robots in highly dynamic and uncertain environments has become a pressing technical challenge for the industry. Summary of the Invention

[0004] This application provides an artificial intelligence-based industrial robot control system and method, which can realize feedforward active collision avoidance control of elastic robots in highly dynamic and uncertain environments.

[0005] In a first aspect, this application provides an artificial intelligence-based industrial robot control method for an industrial robot with multiple switchable drive modes, the method comprising the following steps: A digital twin elastic motion model of the target robot is constructed, and the equivalent elastic parameter distribution of the model is updated by real-time sensor data. Based on the equivalent elastic parameter distribution and the current driving mode of the target robot, the spatiotemporal distribution characteristics of elastic deformation generated during multi-mode switching are predicted using a force input-position output graph neural network. Based on the generalized momentum observer, the force mutation signal of the target robot at the instant of switching between various driving modes is estimated in real time. The non-parametric modeled disturbance torque component is separated from the force mutation signal by the residual reconstructor, and the disturbance propagation path topology is calculated according to the spatiotemporal propagation characteristics of the non-parametric modeled disturbance torque component. The spatiotemporal distribution characteristics of elastic deformation and the topology of the disturbance propagation path are converted into tensors of the same dimension and element-wise multiplied to generate a dynamic risk field. The dynamic risk field is modeled as a Gaussian random field. Collision heatmaps are obtained through random field sampling methods. The collision heatmaps are used as observation variables and input into a counterfactual reasoning model based on causal inference to generate a multi-level reverse planning avoidance strategy tree. The optimal avoidance branch that satisfies the operational constraints is selected from the avoidance strategy tree, and the graded collision avoidance action is executed through the offline pre-trained risk graded response controller.

[0006] In some embodiments, constructing a digital twin elastic motion model of the target robot and updating the equivalent elastic parameter distribution of the model using real-time sensor data specifically includes: The elastic material parameters and geometric parameters of the target robot under multiple driving modes are obtained, and an initial finite element elastic model is established.

[0007] Strain field data is collected in real time by distributed fiber optic strain sensors deployed on key nodes of the robot. The sampling frequency can be 500Hz-2000Hz. Key nodes include joints, link midpoints, end effector mounting bases and other locations with large elastic deformation. At least three orthogonal fiber optic strain sensors are deployed on each key node to collect three-dimensional strain components. Using the strain field data as observations, the local equivalent elastic modulus distribution of the finite element elastic model is inverted and updated to obtain the dynamically updated equivalent elastic parameter distribution. The state vector of the extended Kalman filter is the equivalent elastic modulus of each element in the finite element model, and the observation vector is the strain value collected by each fiber optic sensor. The state vector is updated by minimizing the residual between the observed strain and the model predicted strain.

[0008] In some embodiments, predicting the spatiotemporal distribution characteristics of elastic deformation generated during multi-mode switching using a force input-bit output graph neural network specifically includes: Each rigid body component and elastic connector of the target robot is modeled as a rigid node and an elastic edge in a graph neural network. The features of the rigid node may include node mass, moment of inertia, current position and velocity; the features of the elastic edge include equivalent stiffness, equivalent damping and connection relationship. The driving torque and end load under the current driving mode are used as input features, and the predicted deformation displacement of each node is used as output. By using the forward propagation of the force-input-output graph neural network, the spatial distribution and temporal delay characteristics of elastic deformation transmitted from the driving end to the end during mode switching are calculated, thus obtaining the spatiotemporal distribution characteristics of elastic deformation. This force-input-output graph neural network can employ a structure of 3 graph convolutional layers plus 2 temporal convolutional layers. The graph convolutional layers are used to extract spatial topological features, and the temporal convolutional layers are used to extract temporal evolution features, thereby simultaneously predicting the deformation displacement of each node at different times.

[0009] In some embodiments, the method of estimating the force mutation signal of the target robot at the instant of switching between various driving modes in real time based on a generalized momentum observer, and separating the non-parametric modeled disturbance torque component from the force mutation signal through a residual reconstructor specifically includes: The nonlinear external torque τ_ext = τ_total - τ_dyn of each joint is estimated using a generalized momentum observer, where τ_total is the total driving torque, calculated from the motor current; and τ_dyn is the nominal dynamic torque, calculated from the robot dynamics model. The nonlinear external torque is decomposed into parametric friction torque and nonparametric modeled disturbance torque. The parametric friction torque can be modeled using the Coulomb-viscous friction model, and the model parameters can be calibrated through offline experiments. The nonparametric modeled disturbance torque may include unmodeled nonlinear friction force, external collision force, and load variation force, etc. The time-frequency domain reconstruction of the non-parametric modeled disturbance torque can be performed using sparse Bayesian learning methods to obtain the time-driving mode distribution matrix of the disturbance components.

[0010] In some embodiments, calculating the topology of the disturbance propagation path specifically includes: Based on the time-drive mode distribution matrix of the disturbance component, the starting point of the disturbance in the current drive mode is determined, where the starting point can be the joint node with the largest disturbance energy. Based on the robot kinematic tree structure, the set of possible paths for a disturbance to propagate from the initial action node to adjacent nodes can be calculated using Dijkstra's shortest path algorithm. The robot kinematic tree structure is pre-established, with the root node being the robot base and the leaf nodes being the end effectors. Each path in the set of possible paths is assigned a propagation probability and a propagation delay, forming a weighted directed perturbation propagation path topology. The propagation probability is inversely proportional to the stiffness of each joint on the path, while the propagation delay is directly proportional to the path length and joint damping.

[0011] In some embodiments, converting the spatiotemporal distribution characteristics of elastic deformation and the topology of the disturbance propagation path into tensors of the same dimension and performing element-wise multiplication to generate a dynamic risk field specifically includes: The spatiotemporal distribution characteristics of elastic deformation are represented as a five-dimensional tensor of space-driving mode-time. ,in X, Y, Z For spatial coordinate dimensions, M For the driving mode dimension, T With time as the dimension, the value of each element in the tensor represents the magnitude of the elastic deformation at the corresponding spatial location, driving mode, and time point; The topology of the disturbance propagation path is mapped to a propagation probability tensor P of the same dimension. The value of each element in the tensor represents the probability that the corresponding spatial location, driving mode and time point are affected by the disturbance. Calculate the Hadamard product R = D ⊙ P, and apply it to the driving mode dimension. M Marginalization is performed to obtain a three-dimensional dynamic risk field. Marginalization can be performed by summing, that is, for each spatial-temporal point, the risk values ​​under all driving modes are added together to obtain the total risk value at that point.

[0012] In some embodiments, the dynamic risk field is modeled as a Gaussian random field, and the collision heatmap is obtained through random field sampling methods, specifically including: The dynamic risk field can be viewed as a Gaussian random field, where the risk intensity at each spatial-temporal point follows a Gaussian distribution. The mean μ(x,t) is the definite value of the dynamic risk field at that point, and the variance... Determined by both sensor noise and model error; The Gaussian random field can be sampled 1000 times using the Latin hypercube sampling method to obtain multiple risk intensity samples; Statistical analysis was performed on risk intensity samples at various spatial and temporal points to calculate the collision probability. ,in Generate a collision heatmap based on a preset risk threshold. H(x,t) In the collision heatmap, the color depth of each pixel represents the probability of a collision at the corresponding spatial location at the corresponding time point.

[0013] In some embodiments, inputting the collision heatmap as an observed variable into a counterfactual reasoning model based on causal inference to generate a multi-level reverse planning avoidance strategy tree specifically includes: pre-establishing a structured causal model G = (V, E) between driving modes, control parameters, and collision risks, where V represents variable nodes, including driving mode nodes, joint velocity nodes, trajectory offset nodes, and collision risk nodes; E represents causal edges, indicating the causal relationship between variables; The collision heatmap is used as the observed values ​​of the collision risk variable in the causal graph and input into the structured causal model. Using the absence of a collision as a counterfactual condition, i.e. setting the collision risk variable to 0, the minimum adjustment required for each level of control parameters is calculated in reverse through Judea Pearl's three-step counterfactual reasoning (abduction, action, prediction). By performing multi-level reverse derivation based on time reversal and causal distance, a multi-level reverse planning evasion strategy tree is generated, with the root node as the final evasion action and the leaf nodes as the current mode parameter adjustment. The tree depth is 3-5 levels, and each level corresponds to a parameter adjustment scheme within a time window.

[0014] In some embodiments, selecting the optimal avoidance branch that satisfies the operational constraints from the avoidance strategy tree and executing graded collision avoidance actions through an offline pre-trained risk graded response controller specifically includes: The optimization objectives are to minimize the operation time and execution energy consumption, i.e., min (α·T_execution + β·E_energy), where α and β are weight coefficients that are pre-set according to the operation requirements. The path from the leaf node to the root node that satisfies the joint limit and speed constraint is searched from the avoidance strategy tree and is taken as the optimal avoidance branch. The control parameters at each level on the optimal avoidance branch are arranged in chronological order to form a hierarchical collision avoidance action sequence. For example, the first level is mode switching delay, the second level is motion speed adjustment, and the third level is trajectory offset. The pre-trained neural network controller executes each action in the hierarchical collision avoidance sequence sequentially in a cascaded manner. Upon confirmation of each action's execution, the next level of action is automatically triggered. The pre-trained neural network controller is trained using a deep deterministic policy gradient algorithm, enabling it to generate smooth control commands based on the current robot state and control parameters.

[0015] In some embodiments, the industrial robot is an industrial robot with multiple interchangeable drive modes.

[0016] Secondly, this application provides an artificial intelligence-based industrial robot control system, which includes a collaborative control unit, the collaborative control unit comprising: The deformation prediction module is used to construct a digital twin elastic motion model of the target robot and update the equivalent elastic parameter distribution of the model through real-time sensor data. Based on the equivalent elastic parameter distribution and the current driving mode of the target robot, the module uses a force input-position output graph neural network to predict the spatiotemporal distribution characteristics of elastic deformation generated during multi-mode switching. The disturbance separation module is used to estimate the force mutation signal of the target robot at the moment of switching between various driving modes in real time based on the generalized momentum observer. It separates the non-parametric modeled disturbance torque component from the force mutation signal through the residual reconstructor and calculates the disturbance propagation path topology based on the spatiotemporal propagation characteristics of the non-parametric modeled disturbance torque component. The risk field generation module is used to convert the spatiotemporal distribution characteristics of elastic deformation and the topology of the disturbance propagation path into tensors of the same dimension and perform element-wise multiplication to generate a dynamic risk field. The causal reasoning module is used to model the dynamic risk field as a Gaussian random field, obtain a collision heatmap through a random field sampling method, and input the collision heatmap as an observation variable into the counterfactual reasoning model based on causal inference to generate a multi-level reverse planning avoidance strategy tree. The collision avoidance execution module is used to select the optimal avoidance branch that satisfies the operation constraints from the avoidance strategy tree, and execute the graded collision avoidance action through the offline pre-trained risk graded response controller.

[0017] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: This application utilizes a digital twin model to update the robot's equivalent elastic parameter distribution in real time, combined with graph neural networks to accurately predict the spatiotemporal characteristics of elastic deformation under multi-mode switching. Simultaneously, it leverages a generalized momentum observer and residual reconstruction to achieve accurate perception and propagation topology calculation of unknown disturbances. A dynamic risk field is constructed through tensor fusion, and a collision heatmap is generated probabilistically. Furthermore, multi-level collision avoidance strategies are generated in advance using causal inference and counterfactual reasoning, ultimately achieving precise collision avoidance execution through a hierarchical controller. This solution overcomes the limitations of traditional feedforward and passive collision avoidance control, completing advance prediction and active planning based on real-time dynamic operating conditions throughout the entire process. It effectively adapts to uncertainties such as robot elastic deformation, sudden disturbances, and dynamic environmental changes during operation, stably realizing feedforward active collision avoidance control for elastic robots in highly dynamic and uncertain environments. Attached Figure Description

[0018] Figure 1 is an overall architecture diagram of an artificial intelligence-based industrial robot control system according to some embodiments of this application; Figure 2 is an exemplary flowchart of an artificial intelligence-based industrial robot control method according to some embodiments of this application; Figure 3 This is a schematic diagram of the structure of the collaborative control unit according to some embodiments of this application; Figure 4 is a schematic diagram of the structure of a computer device for implementing an artificial intelligence-based industrial robot control method according to some embodiments of this application. Detailed Implementation

[0019] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0020] refer to Figure 1 The figure is an overall architecture diagram of an artificial intelligence-based industrial robot control system according to some embodiments of this application. This application is divided into two core modules that are coupled with each other: a real robot cluster and a digital twin collaborative control platform. The two interact through three bidirectional links: data acquisition, command issuance and model update.

[0021] The left-hand robot cluster comprises multiple industrial robots with switchable drive modes. Each robot is equipped with a high-precision joint encoder, distributed fiber optic strain sensor, six-dimensional force sensor, momentum observation module, and multi-mode drive actuator. It can collect its own motion state data, elastic deformation data, task execution data, and environmental perception data in real time, and receive control commands from the digital twin collaborative control platform to complete drive mode switching and graded collision avoidance actions. The right-hand digital twin collaborative control platform, with a collaborative control unit at its core, standardizes and integrates multi-source heterogeneous data from the robot cluster through a data fusion module. It relies on a core model module (including a digital twin elastic motion model, force input...) The system comprises a bit-output graph neural network, a generalized momentum observer and residual reconstructor, a dynamic risk field generator, a random field sampling module, a causal inference counterfactual reasoning model, and a risk-level response controller. It enables prediction of the spatiotemporal distribution characteristics of elastic deformation, modeling of disturbance separation and propagation, generation of dynamic risk fields, construction of collision heatmaps, and reasoning of avoidance strategy trees. The decision execution module generates a hierarchical collision avoidance action sequence and sends it to the real robot. Finally, the model update module uses real execution data to synchronously update the parameters of each core model, realizing continuous iterative optimization of the collaborative control platform. This will not be elaborated further here.

[0022] refer to Figure 2 The figure is an exemplary flowchart of an artificial intelligence-based industrial robot control method according to some embodiments of this application. The method mainly includes the following steps: In step 101, a digital twin elastic motion model of the target robot is constructed, and the equivalent elastic parameter distribution of the model is updated by real-time sensor data. Based on the equivalent elastic parameter distribution and the current driving mode of the target robot, the spatiotemporal distribution characteristics of elastic deformation generated during multi-mode switching are predicted using a force input-position output graph neural network.

[0023] Specifically, firstly, the elastic material parameters (e.g., Young's modulus, Poisson's ratio) and geometric parameters (e.g., link length, moment of inertia) of the target robot under multiple driving modes are acquired to establish an initial finite element elastic model. Then, strain field data is collected in real time using distributed fiber optic strain sensors deployed on key nodes of the robot. The sampling frequency can be 500Hz-2000Hz. Key nodes include locations with significant elastic deformation, such as joint connections, link midpoints, and end effector mounting bases. At least three orthogonal fiber optic strain sensors are deployed at each key node to collect three-dimensional strain components. Finally, the strain field data is used as observations, and the local equivalent elastic modulus distribution of the finite element elastic model is updated by inverting, for example, using an extended Kalman filter method, to obtain a dynamically updated equivalent elastic parameter distribution. The state vector of the extended Kalman filter can be the equivalent elastic modulus of each element in the finite element model, and the observation vector can be the strain value collected by each fiber optic sensor. The state vector is updated by minimizing the residual between the observed strain and the model's predicted strain. This is merely an example and is not intended to limit the scope of this application.

[0024] Furthermore, the rigid components and elastic connectors of the target robot are modeled as rigid nodes and elastic edges in a graph neural network. The features of a rigid node include node mass, moment of inertia, current position, and velocity; the features of an elastic edge include equivalent stiffness, equivalent damping, and connectivity. The driving torque and end-effector load in the current driving mode are used as input features, and the predicted deformation displacement of each node is used as output. Through the forward propagation of the force-input-output graph neural network, the spatial distribution and temporal delay characteristics of elastic deformation transmitted from the driving end to the end-effector during mode switching are calculated, resulting in the spatiotemporal distribution characteristics of elastic deformation. It should be noted that, for example, the force-input-output graph neural network in this application can adopt a structure of 3 graph convolutional layers plus 2 temporal convolutional layers. The graph convolutional layers are used to extract spatial topological features, while the temporal convolutional layers are used to extract temporal evolution features, thereby simultaneously predicting the deformation displacement of each node at different times. This is merely an example and is not intended to limit the scope of this application.

[0025] In step 102, the force mutation signal of the target robot at the moment of switching between each driving mode is estimated in real time based on the generalized momentum observer. The non-parametric modeled disturbance torque component is separated from the force mutation signal by the residual reconstructor, and the disturbance propagation path topology is calculated according to the spatiotemporal propagation characteristics of the non-parametric modeled disturbance torque component.

[0026] Specifically, the nonlinear external torque of each joint can first be estimated using a generalized momentum observer: τ_ext = τ_total - τ_dyn, where τ_total is the total driving torque, which can be calculated from the motor current; and τ_dyn is the nominal dynamic torque, which can be calculated from the robot dynamics model. Then, the nonlinear external torque is decomposed into parameterized friction torque and non-parametric modeled disturbance torque. The parameterized friction torque can be modeled using a Coulomb-viscous friction model, and the model parameters can be calibrated through offline experiments; while the non-parametric modeled disturbance torque may include unmodeled nonlinear friction forces, external collision forces, and load change forces, etc. In this application, for example, the non-parametric modeled disturbance torque can be reconstructed in the time-frequency domain using a sparse Bayesian learning method to obtain the time-drive mode distribution matrix of the disturbance component. Since the sparse Bayesian learning method can automatically select the time-frequency basis functions that contribute significantly to the disturbance, it can achieve accurate reconstruction of non-stationary disturbances. Next, the initial action node of the disturbance in the current driving mode is determined based on the matrix. The initial action node is the joint node with the largest disturbance energy. In this application, based on the robot kinematic tree structure, the set of possible paths for the disturbance to propagate from the initial action node to adjacent nodes is calculated using Dijkstra's shortest path algorithm. The robot kinematic tree structure is pre-established, with the root node being the robot base and the leaf nodes being the end effectors. Each path in the set of possible paths is assigned a propagation probability and a propagation delay, forming a weighted directed disturbance propagation path topology. The propagation probability is inversely proportional to the stiffness of each joint on the path, and the propagation delay is directly proportional to the path length and joint damping.

[0027] In step 103, the spatiotemporal distribution characteristics of elastic deformation and the topology of the disturbance propagation path are converted into tensors of the same dimension and element-wise multiplied to generate a dynamic risk field. The dynamic risk field is modeled as a Gaussian random field, and a collision heatmap is obtained through a random field sampling method.

[0028] Specifically, the spatiotemporal distribution characteristics of elastic deformation are represented as a five-dimensional tensor of space (X,Y,Z)-driving mode (M)-time (T). ,in X, Y, Z For spatial coordinate dimensions, M For the driving mode dimension, T The time dimension is used. The value of each element in the tensor represents the magnitude of elastic deformation at the corresponding spatial location, driving mode, and time point. The perturbation propagation path topology is mapped to a propagation probability tensor P of the same dimension. The value of each element in the tensor represents the probability of being affected by the perturbation at the corresponding spatial location, driving mode, and time point. The Hadamard product R = D ⊙ P is calculated, and the driving mode dimension M is marginalized to obtain a three-dimensional dynamic risk field. The marginalization operation uses a summation method, that is, for each spatial-temporal point, the risk values ​​of all driving modes are added together to obtain the total risk value at that point.

[0029] In addition, in practical implementation, the dynamic risk field can be regarded as a Gaussian random field, where the risk intensity at each spatial-temporal point follows a Gaussian distribution. Where the mean μ(x,t) is the definite value of the dynamic risk field at that point, and the variance... The risk intensity is determined by both sensor noise and model error. For example, this application can use the Latin hypercube sampling method to sample the Gaussian random field 1000 times to obtain multiple risk intensity samples. This is because the Latin hypercube sampling method can achieve good spatial coverage with fewer sampling times. The risk intensity samples at each spatial-temporal point are then statistically analyzed to calculate the collision probability. ,in A collision heatmap H(x,t) is generated based on a preset risk threshold. The color depth of each pixel in the collision heatmap represents the probability of collision at the corresponding spatial location at the corresponding time point, which will not be elaborated here.

[0030] In step 104, the collision heatmap is input as an observation variable into the counterfactual reasoning model based on causal inference to generate a multi-level reverse planning avoidance strategy tree.

[0031] Specifically, a structured causal model G = (V, E) can be pre-established, relating driving modes, control parameters, and collision risks. Here, V represents variable nodes, including driving mode nodes, joint velocity nodes, trajectory offset nodes, and collision risk nodes; E represents causal edges, indicating the causal relationships between variables. The collision heatmap is used as the observed values ​​of the collision risk variable in the causal graph and input into the structured causal model. Using the absence of a collision as a counterfactual condition (i.e., setting the collision risk variable to 0), Pearl's three-step counterfactual reasoning (i.e., abduction, action, prediction) is used to calculate the minimum adjustment required for each level of control parameters. Multi-level reverse derivation is performed according to chronological order and causal distance to generate a multi-level reverse planning avoidance strategy tree with the root node representing the final avoidance action and the leaf nodes representing the current mode parameter adjustments. The tree depth can be, for example, 3-5 levels, with each level corresponding to a parameter adjustment scheme within a time window.

[0032] In step 105, the optimal avoidance branch that satisfies the operational constraints is selected from the avoidance strategy tree, and the graded collision avoidance action is executed through the offline pre-trained risk graded response controller.

[0033] Specifically, the optimization objectives are to minimize the operation time and execution energy consumption, i.e., min (α·T_execution + β·E_energy), where α and β are weighting coefficients pre-set according to the operation requirements. The path from the leaf node to the root node that satisfies joint limits and velocity constraints is searched from the avoidance strategy tree and becomes the optimal avoidance branch. The control parameters at each level on the optimal avoidance branch are arranged in chronological order to form a hierarchical collision avoidance action sequence. For example, the first level is mode switching delay, the second level is motion speed adjustment, and the third level is trajectory offset. It should be noted that in this application, a pre-trained neural network controller executes each action in the hierarchical collision avoidance action sequence sequentially in a cascaded manner. After the execution of each action is confirmed, the next level action is automatically triggered. The pre-trained neural network controller is trained using a deep deterministic policy gradient algorithm, which can generate smooth control commands based on the current robot state and control parameters.

[0034] Through the above steps, this application realizes feedforward active collision avoidance control that does not rely on predefined motion primitive sequences and historical execution residual statistics. It is particularly suitable for collaborative operation scenarios of lightweight or long-arm multi-mode industrial robot clusters where elastic deformation cannot be ignored.

[0035] This application also provides an artificial intelligence-based industrial robot control system, which includes a collaborative control unit, as referenced. Figure 3 The figure is a schematic diagram of the structure of a collaborative control unit according to some embodiments of this application. The collaborative control unit includes: a deformation prediction module 301, a disturbance separation module 302, a risk field generation module 303, a causal reasoning module 304, and a collision avoidance execution module 305, which are described below: The deformation prediction module 301 in this application is mainly used to construct a digital twin elastic motion model of the target robot, and update the equivalent elastic parameter distribution of the model through real-time sensor data. Based on the equivalent elastic parameter distribution and the current driving mode of the target robot, the force input-bit output graph neural network is used to predict the spatiotemporal distribution characteristics of elastic deformation generated during multi-mode switching. The disturbance separation module 302 in this application is mainly used to estimate the force change signal of the target robot at the moment of switching between various driving modes based on the generalized momentum observer in real time, separate the non-parametric modeled disturbance torque component from the force change signal through the residual reconstructor, and calculate the disturbance propagation path topology according to the spatiotemporal propagation characteristics of the non-parametric modeled disturbance torque component. Risk field generation module 303: In this application, the risk field generation module 303 is mainly used to convert the spatiotemporal distribution characteristics of elastic deformation and the topology of the disturbance propagation path into tensors of the same dimension and perform element-wise multiplication to generate a dynamic risk field. Causal reasoning module 304: In this application, causal reasoning module 304 is mainly used to model the dynamic risk field as a Gaussian random field, obtain a collision heat map through a random field sampling method, and input the collision heat map as an observation variable into the counterfactual reasoning model based on causal inference to generate a multi-level reverse planning avoidance strategy tree. The collision avoidance execution module 305 in this application is mainly used to select the optimal avoidance branch that satisfies the operation constraints from the avoidance strategy tree, and execute the graded collision avoidance action through the offline pre-trained risk graded response controller.

[0036] In addition, this application also provides a computer device, which includes a memory and a processor. The memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the computer device executes the above-described artificial intelligence-based industrial robot control method.

[0037] In some embodiments, reference Figure 4 The figure is a schematic diagram of the structure of a computer device implementing an artificial intelligence-based industrial robot control method according to some embodiments of this application. The method in the above embodiments can be... Figure 4 The computer device shown is used to implement this, and the computer device 400 includes at least one processor 401, a communication bus 402, a memory 403, and at least one communication interface 404.

[0038] The processor 401 may be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more for controlling the execution of the artificial intelligence-based industrial robot control method of this application.

[0039] The communication bus 402 may include a path for transmitting information between the aforementioned components.

[0040] The memory 403 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, a random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a read-only optical disc (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), a magnetic disk or other magnetic storage device, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 403 may exist independently and be connected to the processor 401 via a communication bus 402. The memory 403 may also be integrated with the processor 401.

[0041] The memory 403 stores program code for executing the scheme of this application, and its execution is controlled by the processor 401. The processor 401 executes the program code stored in the memory 403. The program code may include one or more software modules. In the above embodiments, the construction of the digital twin elastic motion model, the training of the force input-bit output graph neural network, and the reasoning of the counterfactual reasoning model can be implemented by the processor 401 and one or more software modules in the program code in the memory 403.

[0042] Communication interface 404 uses any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, wireless access networks, wireless LANs, etc.

[0043] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core processor or a multi-core processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0044] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, an industrial control computer, an embedded device, etc. This application does not limit the type of computer device.

[0045] In addition, this application also provides a computer-readable storage medium storing instructions or code that, when executed on a computer, cause the computer to perform the aforementioned artificial intelligence-based industrial robot control method.

[0046] Although preferred embodiments of this application 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 the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0047] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. An industrial robot control method based on artificial intelligence, characterized in that, The method includes the following steps: A digital twin elastic motion model of the target robot is constructed, and the equivalent elastic parameter distribution of the model is updated by real-time sensor data. Based on the equivalent elastic parameter distribution and the current driving mode of the target robot, the spatiotemporal distribution characteristics of elastic deformation generated during multi-mode switching are predicted using a force input-position output graph neural network. Based on the generalized momentum observer, the force mutation signal of the target robot at the instant of switching between various driving modes is estimated in real time. The non-parametric modeled disturbance torque component is separated from the force mutation signal by the residual reconstructor, and the disturbance propagation path topology is calculated according to the spatiotemporal propagation characteristics of the non-parametric modeled disturbance torque component. The spatiotemporal distribution characteristics of elastic deformation and the topology of the disturbance propagation path are converted into tensors of the same dimension and element-wise multiplied to generate a dynamic risk field. The dynamic risk field is modeled as a Gaussian random field. Collision heatmaps are obtained through random field sampling methods. The collision heatmaps are used as observation variables and input into a counterfactual reasoning model based on causal inference to generate a multi-level reverse planning avoidance strategy tree. The optimal avoidance branch that satisfies the operational constraints is selected from the avoidance strategy tree, and the graded collision avoidance action is executed through the offline pre-trained risk graded response controller.

2. The method according to claim 1, characterized in that, Constructing a digital twin elastic motion model of the target robot and updating the equivalent elastic parameter distribution of the model using real-time sensor data specifically includes: Obtain the elastic material parameters and geometric parameters of the target robot under multiple driving modes, and establish an initial finite element elastic model; Strain field data is collected in real time by distributed fiber optic strain sensors deployed on key nodes of the robot. Using the strain field data as observations, the local equivalent elastic modulus distribution of the finite element elastic model is updated by inverting the extended Kalman filter method, thereby obtaining the dynamically updated equivalent elastic parameter distribution.

3. The method according to claim 1, characterized in that, The specific characteristics of the spatiotemporal distribution of elastic deformation generated during multi-mode switching, predicted using a force input-bit output graph neural network, include: Each rigid body component and elastic connector of the target robot is modeled as a rigid node and an elastic edge in a graph neural network. The driving torque and end load under the current driving mode are used as input features, and the predicted deformation displacement of each node is used as output. By using the forward propagation of the force input-output graph neural network, the spatial distribution and time delay characteristics of elastic deformation transmitted from the driving end to the end during the mode switching process are calculated, and the spatiotemporal distribution characteristics of elastic deformation are obtained.

4. The method according to claim 1, characterized in that, Based on the real-time estimation of the force mutation signal of the target robot during the switching of various driving modes using a generalized momentum observer, the non-parametric modeled disturbance torque component is separated from the force mutation signal by a residual reconstructor, specifically including: The nonlinear external torques of each joint are estimated using a generalized generalized momentum observer; The nonlinear external torque is decomposed into parametric friction torque and nonparametric modeled disturbance torque; The time-frequency domain reconstruction of the non-parametrically modeled disturbance torque yields the time-driving mode distribution matrix of the disturbance components.

5. The method according to claim 4, characterized in that, The calculation of the disturbance propagation path topology specifically includes: Based on the time-drive mode distribution matrix of the disturbance component, determine the starting point of the disturbance in the current drive mode; Based on the robot's kinematic tree structure, the set of possible paths for a disturbance to propagate from the initial node to adjacent nodes is calculated using the shortest path algorithm in graph theory. Each path in the set of possible paths is assigned a propagation probability and a propagation delay to form a weighted directed perturbation propagation path topology.

6. The method according to claim 1, characterized in that, The process of converting the spatiotemporal distribution characteristics of elastic deformation and the topology of the disturbance propagation path into a tensor of the same dimension and performing element-wise multiplication to generate a dynamic risk field specifically includes: The spatiotemporal distribution characteristics of elastic deformation are represented as a five-dimensional tensor of space-driving mode-time. ,in X, Y, Z For spatial coordinate dimensions, M For the driving mode dimension, T The time dimension; Map the perturbation propagation path topology to a propagation probability tensor P of the same dimension; Calculate the Hadamard product of the five-dimensional tensor and the propagation probability tensor, and then apply the driving mode dimension... M By marginalizing the risk, we obtain a risk intensity field distributed in spatial and temporal dimensions, which serves as a dynamic risk field.

7. The method according to claim 1, characterized in that, The dynamic risk field is modeled as a Gaussian random field, and the collision heatmap is obtained through random field sampling methods, specifically including: The dynamic risk field is viewed as a Gaussian random field, where the risk intensity at each spatial-temporal point follows a Gaussian distribution. Multiple risk intensity samples were obtained by sampling the Gaussian random field multiple times. Statistical analysis is performed on risk intensity samples at each spatial-temporal point to calculate collision probabilities and generate collision heatmaps.

8. The method according to claim 1, characterized in that, The collision heatmap is input as an observed variable into a counterfactual reasoning model based on causal inference to generate a multi-level reverse planning avoidance strategy tree, specifically including: Establish a structured causal model in advance between driving mode, control parameters, and collision risk; The collision heatmap is used as the observed values ​​of the collision risk variable in the causal graph and input into the structured causal model. Using the absence of a collision as a counterfactual condition, the minimum adjustment required for each level of control parameters is calculated through reverse reasoning. By performing multi-level reverse derivation based on time reversal and causal distance, a multi-level reverse planning evasion strategy tree is generated, with the root node as the final evasion action and the leaf nodes as the current mode parameter adjustments.

9. The method according to claim 1, characterized in that, Selecting the optimal avoidance branch that satisfies the job constraints from the avoidance strategy tree and executing graded collision avoidance actions through an offline pre-trained risk graded response controller specifically includes: With the goal of minimizing operation time and execution energy consumption, the path from the leaf node to the root node that satisfies the joint limit and velocity constraints is searched in the avoidance strategy tree and is taken as the optimal avoidance branch. Arrange the control parameters at each level on the optimal avoidance branch in chronological order to form a hierarchical collision avoidance action sequence; The pre-trained neural network controller executes each action in the graded collision avoidance sequence in a cascaded manner.

10. An artificial intelligence-based industrial robot control system, comprising a collaborative control unit, characterized in that, The collaborative control unit includes: The deformation prediction module is used to construct a digital twin elastic motion model of the target robot and update the equivalent elastic parameter distribution of the model through real-time sensor data. Based on the equivalent elastic parameter distribution and the current driving mode of the target robot, the module uses a force input-position output graph neural network to predict the spatiotemporal distribution characteristics of elastic deformation generated during multi-mode switching. The disturbance separation module is used to estimate the force mutation signal of the target robot at the moment of switching between various driving modes in real time based on the generalized momentum observer. It separates the non-parametric modeled disturbance torque component from the force mutation signal through the residual reconstructor and calculates the disturbance propagation path topology based on the spatiotemporal propagation characteristics of the non-parametric modeled disturbance torque component. The risk field generation module is used to convert the spatiotemporal distribution characteristics of elastic deformation and the topology of the disturbance propagation path into tensors of the same dimension and perform element-wise multiplication to generate a dynamic risk field. The causal reasoning module is used to model the dynamic risk field as a Gaussian random field, obtain a collision heatmap through a random field sampling method, and input the collision heatmap as an observation variable into the counterfactual reasoning model based on causal inference to generate a multi-level reverse planning avoidance strategy tree. The collision avoidance execution module is used to select the optimal avoidance branch that satisfies the operation constraints from the avoidance strategy tree, and execute the graded collision avoidance action through the offline pre-trained risk graded response controller.

Images