AI-Adaptive Precision Drive Control Optimization Method for Robot Equipment
By combining online clustering confidence assessment and pre-trained control strategies, the robot drive system can actively adapt to rapidly changing operating conditions, solving the lag problem of traditional PID control and improving tracking accuracy and operational stability.
Patent Information
- Application Number
- CN202611500306.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing robot drive systems suffer from adjustment lag when operating conditions change rapidly, resulting in decreased tracking accuracy. Traditional PID control is difficult to adapt to frequent switching of operating conditions in high-speed and high-acceleration/deceleration scenarios.
Operating conditions are identified through online clustering confidence assessment, pre-trained control strategy parameters are invoked, and a confidence-weighted exponential smooth transition mechanism is used to gradually switch control parameters. Combined with monitoring of current peak and tracking error during the transition period, the parameters are actively pre-adapted.
It effectively improves the tracking accuracy and operational stability of the robot under high-speed and high-acceleration/deceleration conditions, and avoids current surges and mechanical oscillations caused by parameter adjustment lag.
Smart Images

Figure CN123113383A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of robot drive control technology, specifically to an AI-adaptive precision drive control optimization method for robot equipment. Background Technology
[0002] In high-speed, high-acceleration-deceleration and other variable operating conditions, the control parameters of the drive system of robotic equipment need to be adjusted in real time according to the changes in operating conditions to ensure tracking accuracy and operational stability. Traditional fixed-parameter PID control is difficult to adapt to frequent changes in operating conditions, while most existing adaptive control methods adopt a passive response mode of "adjusting control parameters only after detecting an increase in tracking error", which has the problem of adjustment lag when operating conditions change rapidly.
[0003] In the prior art, invention patent publication number CN121806499A discloses an intelligent optimization method for control parameters of a mobile robot servo motor system. This method distinguishes between steady-state, transitional, and extreme disturbance conditions through working condition evolution identification, establishes a mechanistic mapping model between control parameters and dynamic response indicators, and constructs a multi-objective parameter optimization space. The objective weights are dynamically adjusted based on the working condition identification results. After stabilization and convergence correction of the online optimization results, the target control parameter combination is output. Although this method introduces working condition identification, its core remains an online optimization mode. After working condition identification, real-time multi-objective parameter optimization calculations are required. There is a non-negligible time delay between the occurrence of working condition changes, optimization convergence, and parameter loading completion, which is a typical passive response path of identification, optimization, and adjustment. Another invention patent publication number CN120447346A discloses an adaptive adjustment method and system for parallel robots based on an attention mechanism. This method obtains a trajectory time-series sample dataset based on real-time trajectory error and joint space control torque, constructs a SeNet-Transformer action network model based on PPO reinforcement learning for training, obtains the optimal PID adaptive adjustment model, and outputs the PID adjustment parameters. Although this method uses reinforcement learning for parameter optimization, it still follows the reactive logic of real-time trajectory error input to model inference output parameters, that is, new control parameters are calculated by the model only after the error occurs.
[0004] The aforementioned existing technologies all belong to a passive response mode of changing operating conditions, increasing errors, detection errors, and adjusting parameters. When operating conditions change rapidly, there is an adjustment lag, which leads to a decrease in tracking accuracy in high-speed and high-acceleration / deceleration scenarios. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an AI-adaptive precision drive control optimization method for robot equipment. By using online clustering confidence assessment to reliably identify operating conditions and call the corresponding pre-trained control strategy parameters, and with the help of a confidence-weighted exponential smooth transition mechanism, the control parameters are gradually switched to the target parameters. This achieves proactive pre-adaptation of control parameters before changes in operating conditions and solves the problem of adjustment lag in existing passive response modes when operating conditions change rapidly.
[0006] To address the aforementioned technical problems, this invention provides the following technical solution: an AI-adaptive precision drive control optimization method for robot equipment, the method comprising the following steps: S1: Real-time acquisition of robot joint angular velocity, angular acceleration, load torque, d-axis current, q-axis current and current harmonic amplitude to construct the current working condition feature vector; S2, based on the operating condition feature vector, uses the online density peak clustering algorithm to cluster the current operating conditions, and obtains the operating condition category to which the current operating condition belongs and the Mahalanobis distance of the current operating condition feature vector relative to the center of the operating condition category; S3, calculate the confidence level of the current working condition identification based on the mean and standard deviation of the Mahalanobis distance and the historical samples of the corresponding working condition category; S4: Pre-trained control strategy parameters corresponding to each working condition category are stored in advance. When the confidence level is greater than the first threshold, the pre-trained control strategy parameter corresponding to the current working condition category is called as the target control parameter. When the confidence level is greater than the second threshold and less than or equal to the first threshold, the pre-trained control strategy parameter corresponding to the current working condition category is called and corrected by combining real-time tracking error as the target control parameter. When the confidence level is less than or equal to the second threshold, the current control parameter remains unchanged. S5, exponentially smooth the transition from the current control parameter to the target control parameter according to the following recursive formula: in, This is the control parameter vector at the current moment. This is the control parameter vector for the next control cycle. Let λ be the target control parameter vector, λ be the preset transient rate coefficient, C be the confidence level, and Δt be the control period. The control parameter vector includes a speed loop proportional coefficient, a speed loop integral coefficient, a speed loop differential coefficient, a position loop proportional coefficient, a position loop integral coefficient, a position loop differential coefficient, a speed feedforward coefficient, an acceleration feedforward coefficient, a notch filter center frequency, and a notch filter depth. Through S5, the current control parameters are gradually changed towards the target control parameters to avoid current surges and mechanical oscillations caused by parameter jumps.
[0007] Furthermore, the online density peak clustering algorithm employs a time decay factor and a sliding window mechanism for online updates; The operating condition feature vector is an 8-dimensional feature vector, consisting of angular velocity, angular acceleration, load torque, q-axis current, d-axis current, second harmonic amplitude, fourth harmonic amplitude, and sixth harmonic amplitude. The feature values of each dimension are standardized by Z-score before being used in clustering calculations; Standardization allows operating condition characteristics with different dimensions to participate in cluster distance calculation on a unified scale, eliminating the impact of dimensional differences on clustering results.
[0008] Furthermore, the confidence level is calculated according to the following formula: Where C is the confidence level, and its value ranges from (0,1]. The Mahalanobis distance from the current working condition feature vector to the center of its corresponding working condition category is given. The mean Mahalanobis distance of the historical samples for the corresponding working condition category is updated online using an exponentially weighted moving average. The Mahalanobis distance standard deviation of the historical samples for the corresponding working condition category is updated online using an exponentially weighted moving average. The confidence calculation formula ensures that the greater the deviation of the working condition sample from the cluster center, the lower the confidence value, thus providing a quantitative basis for the classification reliability of subsequent strategy invocation.
[0009] Furthermore, the pre-trained control policy parameters are obtained in the following manner: For each work condition category, a flexible motion-evaluation reinforcement learning algorithm is used for offline training in the robot dynamics simulation environment. The reward function is a weighted negative value of the squared tracking error, the squared rate of change of control quantity, and energy consumption. After training convergence, a 9-dimensional control parameter vector is output as the pre-trained control strategy parameters corresponding to that work condition category. The 9-dimensional control parameter vector includes the velocity loop proportional coefficient, velocity loop integral coefficient, velocity loop differential coefficient, position loop proportional coefficient, position loop integral coefficient, position loop differential coefficient, velocity feedforward coefficient, acceleration feedforward coefficient, and notch filter center frequency and notch filter depth. Offline pre-training ensures that each operating condition category has corresponding convergent optimal control strategy parameters, which can be directly called during online runtime, avoiding the convergence waiting time of online optimization.
[0010] Furthermore, the first threshold is 0.85, and the second threshold is 0.50; When the confidence level is greater than 0.85, the pre-trained control policy parameters are directly used as the target control parameters. When the confidence level is greater than 0.50 and less than or equal to 0.85, the pre-trained control policy parameters are corrected according to the following formula and then used as the target control parameters: in, This is the corrected target control parameter vector. Here, γ is the pre-trained control strategy parameter vector, e(t) is the tracking error at the current time, and G is the pre-set error-parameter gain matrix. By modifying the formula to bring the confidence level to the middle range, the fine-tuning amount of real-time error feedback is superimposed on the pre-trained parameters to compensate for the residual deviation between the pre-trained strategy and the current actual working conditions.
[0011] Furthermore, in S5, transition safety monitoring is also performed: during the transition of control parameters, the peak value of the drive current and the speed tracking error are monitored in real time; when the peak value of the drive current exceeds a preset current threshold or the speed tracking error exceeds a preset error threshold, the current transition is paused and a reverse transition is performed according to the following recursive formula: in, Control parameter vector at the moment before transition start This is the preset reverse transition rate coefficient; By using reverse transition, the parameter transition process can be stopped in time when an abnormality occurs and the system can be rolled back to the safe parameter state before the transition, thus ensuring the safe operation of the drive system during parameter switching.
[0012] Furthermore, S2 also includes the creation of a new working condition category: if the weighted Euclidean distance from the current working condition feature vector to the center of all existing working condition categories is greater than the preset new cluster creation distance threshold, a new working condition category is created with the current working condition feature vector as the initial center; If the number of accumulated samples for a new working condition category reaches the preset minimum number of training samples, the training of the pre-trained control strategy parameters for that new working condition category will be started in the background. After the training is completed, the new working condition category and its corresponding pre-trained control strategy parameters will be added to the pre-stored strategy library. By creating new operating condition categories, unknown operating condition patterns that occur during online operation can be automatically identified and supplemented with corresponding pre-trained strategies, thereby enabling online expansion of the strategy library.
[0013] Furthermore, the method also includes S6: For each preset evaluation cycle, the usage frequency of the pre-trained control strategy parameters corresponding to each working condition category and the average tracking error during the call period are statistically analyzed. Pre-trained control strategy parameters that are used less frequently than the first frequency threshold and whose average tracking error is worse than a preset multiple of the global average tracking error are marked as to be eliminated. After being in a pending elimination state for several consecutive evaluation cycles, the operating condition category and its corresponding pre-trained control strategy parameters are deleted from the pre-stored strategy library. By maintaining the policy library, policies that have not been used for a long time or whose control effect is consistently inferior to the global level are automatically eliminated, thus controlling the size of the policy library and ensuring the efficiency of policy query and invocation.
[0014] Furthermore, S2 also includes merging and deleting operating condition categories: When the distance between the centers of two existing working condition categories is less than the preset merging distance threshold, the two working condition categories are merged into one working condition category, and the center of the merged category is the weighted average of the centers of the two working condition categories. If no new samples are added to any work condition category within a preset time period, delete that work condition category and its corresponding pre-trained control strategy parameters. By merging and deleting, the number of working condition categories is kept compact, avoiding increased computational overhead for strategy queries due to redundancy in working condition categories.
[0015] Furthermore, the method also includes S7: Store the feature vectors of operating conditions with a confidence level less than or equal to the second threshold into the cache area; Each time a preset batch update cycle is run, batch re-clustering is performed on the working condition feature vectors in the cache area, and the newly formed dense region after re-clustering is re-matched with the existing working condition categories or a new working condition category is created. The cluster centers in the online density peak clustering algorithm are updated using an exponentially weighted moving average method. in, This is the updated cluster center vector. t is the cluster center vector before the update, β is the preset cluster center update decay coefficient, and F(t) is the working condition feature vector at the current time. By using exponentially weighted moving average updates, the center of the operating condition category evolves smoothly and continuously as operating data accumulates, tracking the long-term, slow changes in operating condition characteristics and maintaining the timeliness of the clustering model.
[0016] Compared with existing technologies, this AI-adaptive robot equipment precision drive control optimization method has the following advantages: I. This invention uses online clustering confidence assessment to identify and quantify the reliability of robot operating conditions in real time. When the operating conditions change, it directly calls the pre-trained control strategy parameters as target parameters, so that the control parameters can respond to changes in operating conditions in advance rather than passively adjusting after errors occur. This eliminates the parameter adjustment lag in scenarios with rapid changes in operating conditions and effectively improves the tracking accuracy under high-speed and high-acceleration / deceleration operating conditions.
[0017] Second, this invention uses a confidence-weighted exponential smoothing recursive formula to gradually transition the current control parameters to the target control parameters, making the parameter switching process continuous and smooth. Combined with the current peak and tracking error safety monitoring mechanism during the transition period, it effectively suppresses current surges and mechanical oscillations during parameter switching, and improves the operational stability and safety of the drive system during parameter adaptive adjustment.
[0018] Other advantages, objectives and features of the invention will be set forth in any degree in the description which follows, and in any degree will be apparent to those skilled in the art based on an examination of the following, or may be learned from practice of the invention. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0020] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a detailed flowchart of the clustering and confidence assessment module of the present invention; Figure 3 This is a detailed flowchart of the strategy invocation and parameter transition module of the present invention. Detailed Implementation
[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0022] like Figure 1 As shown, this method mainly includes a data acquisition and preprocessing stage, an online operating condition clustering and confidence assessment stage, a policy pre-adaptation invocation stage, and a control parameter smooth transition stage. The entire process is executed cyclically with a control cycle of 1 kHz, that is, each control cycle... .
[0023] S1, Data Acquisition and Preprocessing: In this embodiment, a six-joint industrial robot is used as the application object, with each joint driven by a permanent magnet synchronous motor and a harmonic reducer. The robot's controller uses an embedded control board based on an ARM Cortex-A53 processor core, with a main frequency of 1.2GHz and equipped with a floating-point unit. The servo drivers of each joint interact with the main controller via an EtherCAT industrial Ethernet bus, with a communication cycle set to 1ms.
[0024] At the start of each control cycle, the main controller reads real-time operating data from each joint servo driver via the EtherCAT bus. The read data includes: joint angular velocity, obtained by differential calculation from an incremental encoder mounted at the motor tail end, with a resolution of 16384 pulses per revolution; joint angular acceleration, calculated from the angular velocity data using a backward differential method, with a differential interval of two consecutive control cycles; estimated load torque, calculated in real-time by the torque observer inside the servo driver based on the q-axis current and torque constant; d-axis and q-axis currents, sampled by current sensors inside the servo driver, with a sampling accuracy of 12 bits and a full-scale range of ±20A; and current harmonic amplitude, extracted by performing a fast Fourier transform on the three-phase current.
[0025] For the extraction of current harmonic amplitudes, this embodiment employs a sliding window Fast Fourier Transform (FFT) method. In each control cycle, the three-phase current data from the most recent 256 sampling points are used to form a data window, with a window length of 256 sampling points corresponding to a 256ms time span. After Hanning window weighting of the current data within this window, a 256-point FFT is performed, and the amplitudes of the 2nd, 4th, and 6th harmonics are extracted from the transform results. The 2nd, 4th, and 6th harmonics are chosen because in a permanent magnet synchronous motor driving a harmonic reducer system, the 2nd harmonic of the motor current mainly reflects the load non-uniformity, the 4th harmonic mainly reflects the non-ideal distribution of the motor rotor magnetic field, and the 6th harmonic mainly reflects the nonlinear characteristics of the inverter switching transistors. These three harmonic components can comprehensively reflect the operating state of the drive system.
[0026] After data acquisition, data preprocessing is performed. Preprocessing consists of two steps: First, a second-order low-pass filter is used to filter the acquired raw signal. The filter cutoff frequency is set to 100Hz to remove high-frequency noise components above 100Hz, primarily originating from the inverter's switching frequency and its harmonics. Second, Z-score standardization is performed on the filtered data for each dimension. Standardization aims to convert physical quantities with different dimensions to a unified scale, preventing unreasonable bias in subsequent clustering distance calculations due to dimensional differences. The standardization formula is: the standardized value of a dimension equals the original value of that dimension minus its historical mean, divided by its historical standard deviation. The historical mean and standard deviation are updated online using an exponentially weighted moving average with a decay coefficient of 0.95, allowing the statistics to adaptively adjust to slow changes in operating conditions.
[0027] After the above data acquisition and preprocessing, an 8-dimensional operating condition feature vector F(t) is generated for each control cycle. This feature vector consists of the following eight components in a fixed order: the first component is angular velocity, the second component is angular acceleration, the third component is load torque, the fourth component is q-axis current, the fifth component is d-axis current, the sixth component is the amplitude of the second current harmonic, the seventh component is the amplitude of the fourth current harmonic, and the eighth component is the amplitude of the sixth current harmonic. This 8-dimensional feature vector will serve as the input data for subsequent online clustering algorithms.
[0028] S2, Online Working Condition Clustering: like Figure 2 As shown, this embodiment employs an improved online density peak clustering algorithm to perform real-time clustering of operating condition feature vectors. Based on the traditional density peak clustering algorithm, this algorithm introduces a time decay factor and a sliding window mechanism, enabling it to adapt to the online processing requirements of data streams.
[0029] Clustering initialization phase: During the first 200 control cycles (the first 200ms) after system power-on, the clustering model is not updated; instead, the feature vectors of the first 200 operating conditions are stored in the initialization buffer. After accumulating 200 samples, the K-means++ algorithm is used to perform offline clustering on these samples, with the initial number of clusters set to 4. The core idea of the K-means++ algorithm is to maximize the distance between the initial cluster centers when selecting them, thereby avoiding getting trapped in local optima. After initialization, four initial cluster centers and initial statistics for each cluster are obtained, including the covariance matrix, mean Mahalanobis distance, and standard deviation of the samples within each cluster.
[0030] Weighted Euclidean distance calculation and sample inclusion: Starting from the 201st control cycle, a clustering update operation is performed once per control cycle. For the current time t, the working condition feature vector F(t) is first calculated to its weighted Euclidean distance to all existing cluster centers. The weighted Euclidean distance is calculated as follows: the square root of the sum of the squared distances in each dimension multiplied by the weight coefficient of that dimension, is given by: in, For weighted Euclidean distance; The i-th dimension feature weight coefficient; The value of the i-th dimension of the current feature vector; This represents the i-th dimension value corresponding to the cluster center.
[0031] The specific calculation methods for the weight coefficients of each dimension are as follows: In each control cycle, the variance of each dimension's eigenvalues is calculated online using an exponentially weighted moving average method. Let the eigenvalue sequence of the i-th dimension be... The mathematical expressions for its exponentially weighted moving average and exponentially weighted moving average variance are: Among them, time decay factor .
[0032] The weight coefficient for each dimension is the ratio of the reciprocal of the variance of that dimension to the sum of the reciprocals of the variances of all dimensions. The mathematical expression is: The weighting coefficient satisfies This ensures that the sum of the weights of each dimension is 1. Dimensions with larger variances have smaller weights, and dimensions with smaller variances have larger weights. Its purpose is to make the clustering algorithm pay more attention to feature dimensions that are stable and have a large amount of information.
[0033] After calculating the weighted Euclidean distances from F(t) to all cluster centers, the cluster with the smallest distance is found, and F(t) is temporarily assigned to that cluster. Then, it is determined whether this smallest distance is greater than a preset threshold for creating new clusters. . The initial value is set to 3 times the average distance within each cluster during the initialization phase, and is continuously updated during subsequent operation.
[0034] If minimum distance This indicates that the current working condition feature vector differs significantly from all existing working condition categories, necessitating the creation of a new working condition category. After the new cluster is created, until the corresponding pre-trained strategy is complete, no strategy switching is performed on samples within that cluster; the current control parameters remain unchanged, and only sample data continues to accumulate. Only after the background training is complete and the new strategy is added to the strategy library can subsequent samples falling into that cluster trigger strategy invocation.
[0035] If minimum distance If so, F(t) will be formally assigned to the nearest cluster.
[0036] Online updates of cluster centers and statistics: For samples assigned to a particular cluster, the cluster centers are updated using an exponentially weighted moving average. The mathematical expression for the update formula is: in, This is the updated cluster center vector. The value is the cluster center vector before the update, and the cluster center update decay coefficient β = 0.97. This value ensures that the cluster center update can follow the slow changes in operating conditions without being disturbed by occasional noise samples.
[0037] For samples assigned to a particular cluster, the statistics for that cluster also need to be updated. This includes the cluster's covariance matrix. Online incremental updates are performed using a recursive covariance update formula, eliminating the need to retain all historical samples. The mathematical expression for the recursive covariance update is as follows: Let the current number of samples in the cluster be n, the cluster center be C, the new sample be F(t), and the updated cluster center be... The updated covariance matrix is: After the update is completed, the number of samples n increases by 1.
[0038] The inverse of the covariance matrix The Sherman-Morrison formula is used for recursive updates, avoiding the need to perform matrix inversion operations each time, thus reducing computational complexity. After obtaining the inverse of the covariance matrix, the Mahalanobis distance from F(t) to the cluster center is calculated, mathematically expressed as: Then, the Mahalanobis distance mean of the cluster is updated using an exponentially weighted moving average. and standard deviation The mathematical expression for mean update is: The standard deviation is updated by first calculating the exponentially weighted moving average of the squared Mahalanobis distance: The standard deviation is then calculated using the following formula: Among them, time decay factor The time decay factor aims to gradually reduce the contribution of older samples to the current statistic, thereby enabling the statistic to reflect the statistical characteristics of recent operating conditions.
[0039] Sliding window mechanism: The sliding window mechanism is implemented as follows: A fixed-length sample queue of 500 samples is maintained in each cluster, corresponding to 0.5 seconds of data. When a new sample is added to a cluster, it is added to the tail of the queue; if the queue length exceeds 500, the oldest sample is removed from the head of the queue. Cluster center updates and statistical updates are performed only based on the 500 samples in the queue; samples outside the queue are not included in the calculation. The 500-sample queue length is based on the fact that 500 samples correspond to 0.5 seconds of running data, containing sufficient statistical information to reflect the operating conditions without reducing the clustering model's response speed to changes in operating conditions due to an excessively long window. The sliding window mechanism, combined with the time decay factor, enables the clustering model to respond quickly to changes in operating conditions while maintaining a certain level of stability.
[0040] Merging of operating condition categories: The merging operation for operating conditions is performed once every 1000 control cycles (1 second). It checks the distance between the cluster centers of all existing clusters; if the distance between two cluster centers is less than a preset merging distance threshold... If the two clusters are not found to be in a given cluster, they will be merged into one. The center of the merged cluster will be the weighted average of the centers of the two clusters, with the weights determined by the number of samples in each cluster; the cluster with more samples will have a larger weight. A merge distance threshold will also be applied. The value is set to 0.5 times the average distance between clusters during the initialization phase. The purpose of the merging operation is to prevent the generation of too many fragmented clusters due to continuous changes in operating conditions, and to maintain the compactness of the number of operating condition categories.
[0041] Deletion of operating condition categories: The deletion of operating condition categories is performed in each control cycle. The most recent sample inclusion time of all existing clusters is checked. If no new samples are included in a cluster within 2000 consecutive control cycles (2 seconds), the operating condition is determined to have disappeared, and the cluster and all its corresponding data, including cluster center, statistics, sample queue, and corresponding pre-trained control policy parameters, are deleted. The purpose of the deletion operation is to promptly clean up invalid operating condition categories and prevent the policy library from expanding indefinitely.
[0042] S3, Confidence calculation: Confidence calculation is a core component of this invention, and its purpose is to provide a quantitative basis for the reliability of working condition identification for subsequent strategy invocation.
[0043] After completing the clustering and incorporation operation in each control cycle, calculate the Mahalanobis distance from the current operating condition feature vector F(t) to the center of its cluster. The Mahalanobis distance is calculated using the inverse of the covariance matrix that has already been updated in S2. The calculation formula is .
[0044] After obtaining the Mahalanobis distance Then, the mean Mahalanobis distance of the historical samples of this cluster was combined. and standard deviation The confidence level C is calculated using the following mathematical formula: The physical meaning of this formula is: when the Mahalanobis distance of the current sample... The average Mahalanobis distance close to the historical samples of this cluster hour, A smaller value and a confidence level C close to 1 indicate that the current operating condition is highly consistent with the typical operating condition of this cluster, and the classification result is reliable; when Gradually increase and exceed When the range is , A significant increase in confidence level C rapidly approaches 0, indicating that the current operating conditions deviate significantly from the typical operating conditions of this cluster, and the classification results are unreliable. As a key term in the denominator, it acts as a normalization scaling factor, making the confidence levels between different clusters comparable. From a statistical perspective, This roughly corresponds to the 95% confidence upper limit of the Mahalanobis distance distribution of the cluster samples. Samples exceeding this range are considered outliers or boundary samples.
[0045] The confidence level C ranges from (0,1). When F(t) is exactly equal to the cluster center, C=1; when As it approaches infinity, C approaches 0. In actual operation, the typical distribution of confidence level C is as follows: under normal operating conditions, C is usually above 0.8; when the operating conditions fluctuate slightly, C is between 0.5 and 0.8; and when the operating conditions change drastically or abnormalities occur, C is below 0.5.
[0046] S4, Building the pre-trained policy library: The pre-trained policy library was built offline. For each work condition category, the SoftActor-Critic reinforcement learning algorithm was used for training in a robot dynamics simulation environment.
[0047] The simulation environment is built based on a robot dynamics simulation platform. Taking the six-joint industrial robot in this embodiment as an example, a high-fidelity dynamic model of the robot is established in the simulation platform. The dynamic model includes: the moment of inertia, reduction ratio, transmission stiffness, and damping coefficient of each joint; the friction model of each joint adopts the LuGre friction model, including parameters such as static friction coefficient, Coulomb friction coefficient, viscous friction coefficient, and Stribeck velocity; the motor model adopts the dq-axis mathematical model of a permanent magnet synchronous motor, including parameters such as stator resistance, d-axis inductance, q-axis inductance, and permanent magnet flux linkage. The control cycle of the simulation environment is set to 1ms, consistent with the physical controller.
[0048] The friction model parameters in the simulation environment were obtained through standard friction force identification experiments on a real robot. Specifically, without a load installed at the robot's end effector, each joint was driven at several different constant speeds between 0.001 rad / s and 1 rad / s. The steady-state driving torque at each speed was recorded, and the parameters of the LuGre model were obtained through nonlinear least squares fitting. If the friction characteristics change due to wear during actual use, the online clustering and confidence evaluation modules automatically trigger policy fine-tuning or new policy training for compensation, without the need to repeat the offline identification experiments.
[0049] During training, for each work condition category, typical distribution ranges of velocity and load are extracted from the sample data of that category, and corresponding training tasks are generated in the simulation environment. The training task takes the form of: given an initial state and a target trajectory, the robot needs to track and run along that target trajectory. The target trajectory can be a sinusoidal trajectory, an S-shaped trajectory, or a spline trajectory, and the maximum velocity, maximum acceleration, and load conditions of the trajectory are determined based on typical parameters of the work condition category.
[0050] The state space of reinforcement learning includes: current joint angle, angular velocity, angular acceleration, tracking error, error integral, and error derivative. The action space is a 9-dimensional control parameter vector, including velocity loop proportional coefficient, velocity loop integral coefficient, velocity loop derivative coefficient, position loop proportional coefficient, position loop integral coefficient, position loop derivative coefficient, velocity feedforward coefficient, acceleration feedforward coefficient, notch filter center frequency, and notch filter depth. Each dimension of the action space is normalized, with values ranging from [0,1], and then mapped to their respective physical value ranges in practical applications.
[0051] The mathematical expression for the reward function is: Where e is the tracking error, Δu is the rate of change of the control quantity, E is the energy consumption, and the weighting coefficient is... , , The meaning of the reward function is: the immediate reward obtained by the system during operation is negative. The larger the tracking error, the more drastic the change in control quantity, and the higher the energy consumption, the greater the negative value of the reward, which means the greater the penalty. The goal of the agent is to maximize the cumulative reward in a round, that is, to minimize the weighted sum of tracking error, change in control quantity, and energy consumption.
[0052] The hyperparameter settings for the SoftActor-Critic algorithm are as follows: Both the Actor and Critic networks are three-layer fully connected networks with 128 neurons per layer and ReLU activation function; the optimizer used is Adam with a learning rate of 3×10⁻⁶. -4 The discount factor γ is 0.99; the target network update rate τ is 0.005; the entropy regularization coefficient α is automatically adjusted, with an initial value of 0.2; and the experience replay pool capacity is 10. 6 The sample size per batch is 256; the number of training rounds is 5000.
[0053] The criterion for training convergence is as follows: Take the average reward value of the most recent 100 training rounds. If the fluctuation range of the average reward value within 50 consecutive rounds does not exceed 2% of the total average reward value of the most recent 100 rounds, then convergence is determined and training is stopped. If the above condition is not met after more than 10,000 training rounds, the policy parameter with the highest cumulative reward among the 10,000 rounds is taken as the final output. The mathematical criterion is: Let the average reward of the most recent 100 rounds be... The average reward per round for 50 consecutive rounds is ,satisfy It converges after 50 rounds.
[0054] Training was performed on a compute server equipped with an NVIDIA RTX 4090 GPU, with each work condition category taking approximately 2 to 4 hours to train. After training converged, the 9-dimensional control parameter vector obtained from the training, along with the cluster center coordinates for that work condition category, was stored in the policy library. The policy library was stored using a hash table structure, with the key being the hash value of the cluster center coordinates and the value being the corresponding control parameter vector. During a query, the hash value was calculated based on the cluster center coordinates of the current sample, and a lookup was performed in the hash table with an O(1) time complexity.
[0055] S5, Policy Pre-Adaptation Invocation: After the confidence level is calculated in each control cycle, different policy invocation logic is executed based on the confidence level value.
[0056] In this embodiment, the first threshold is set to 0.85, and the second threshold is set to 0.50. The selection criteria for these two thresholds are as follows: The first threshold of 0.85 corresponds to the Mahalanobis distance. Approximately equal to The location of the sample within this range indicates that it belongs to the cluster core region, and the classification reliability is very high; the second threshold of 0.50 corresponds to the Mahalanobis distance. Approximately equal to The sample is located within a relatively reliable region of the cluster, but it has begun to deviate from the core.
[0057] When the confidence level C > 0.85, the current operating condition identification is determined to be highly reliable, and the pre-trained control policy parameters corresponding to the current cluster are directly queried from the policy library. This parameter is used as the target control parameter.
[0058] When the confidence level is 0.50, the current operating condition identification is considered relatively reliable but still has some deviation. At this point, the pre-trained control strategy parameters are first queried from the strategy library. Then, the corrected target control parameters are calculated according to the following formula, the mathematical expression of which is: Wherein, the fine-tuning coefficient γ=0.03; e(t) is the position loop tracking error at the current time, in rad; G is a 9×1 dimensional error-parameter gain matrix column vector, corresponding to the error response sensitivity of the 9 control parameters.
[0059] The method for obtaining the error-parameter gain matrix G is as follows: In the simulation environment, set the current control parameters to... The robot is made to track a standard sinusoidal trajectory. After the system enters steady state, a +5% step disturbance is applied to each control parameter, and the tracking error response is continuously observed over 50 control cycles. The ratio of the steady-state error change to the parameter change is taken as the gain value corresponding to that parameter. The gain values obtained for each of the nine parameters are arranged as column vectors to form G. In this embodiment, the mathematical expression of the gain matrix is: The physical meaning of this correction formula is: when the tracking error is positive, a correction amount proportional to the error is added to the pre-trained parameters; when the tracking error is negative, a correction amount proportional to the absolute value of the error is subtracted from the pre-trained parameters. The magnitude of the correction amount is determined by the fine-tuning coefficient. Together with the gain matrix G, it controls the adjustment to ensure that the adjustment magnitude is appropriate and will not compromise the stability of the pre-trained strategy due to excessive adjustment.
[0060] When the confidence level C ≤ 0.50, the current operating condition identification is deemed unreliable, and no strategy switching operation is performed, maintaining the current control parameters unchanged. Simultaneously, this sample is marked as a low-confidence sample and stored in the cache for subsequent incremental updates of the clustering model.
[0061] S6, Smooth transition of control parameters and transition safety monitoring: After determining the target control parameters Afterwards, the current control parameters need to be smoothly transitioned to the target control parameters in order to avoid current surges and mechanical oscillations caused by parameter abrupt changes.
[0062] This embodiment uses a confidence-weighted exponential smoothing recursive formula to achieve parameter transition, the mathematical expression of which is: in, This is the control parameter vector at the current moment. For the control parameter vector of the next control cycle, the transient rate coefficient C represents the confidence level of the current operating condition.
[0063] The value of the transition rate coefficient λ determines the speed of parameter transition. In this embodiment, λ = 0.15. When the confidence level C = 0.9, the parameter moves closer to the target value by 13.5% (0.15 × 0.9) per control cycle, and 90% of the transition can be completed in about 5 ms, or 5 control cycles. When the confidence level C = 0.6, the parameter moves closer to the target value by 9% (0.15 × 0.6) per control cycle, and 90% of the transition can be completed in about 10 ms. This mechanism ensures that parameter switching is fast when the confidence level is high and slow when the confidence level is low, and the transition process remains continuous and smooth, without any abrupt changes.
[0064] In the above exponential smoothing recursive formula, the nine control parameters undergo independent smoothing transitions. Each parameter uses the same λ and C values, but each has its own independent characteristics. and Values. Nine parameters are synchronized during the transition, ensuring that the parameter matching relationship of each part of the controller remains consistent throughout the transition process.
[0065] The specific structure of the control parameter vector is as follows: The first component is the velocity loop proportional coefficient. The value ranges from 0.1 to 100; the second component is the velocity loop integral coefficient. The value ranges from 0.01 to 10; the third component is the velocity loop differential coefficient. The value ranges from 0 to 1; the fourth component is the position ring scaling factor. The value ranges from 1 to 500; the 5th component is the position loop integral coefficient. The value ranges from 0.01 to 50; the 6th component is the differential coefficient of the position ring. The value ranges from 0 to 5; the 7th component is the velocity feedforward coefficient. The value ranges from 0 to 1; the 8th component is the acceleration feedforward coefficient. The value ranges from 0 to 1; the 9th component is the center frequency of the notch filter. and notch filter depth The center frequency of the notch filter is represented by a combination of two parameters, but occupies one dimension in the parameter vector.
[0066] During parameter transitions, this embodiment also implements a transition safety monitoring mechanism. In each control cycle, the peak drive current and speed tracking error are monitored in real time. The peak drive current is obtained by reading the peak current register inside the servo driver, which records the instantaneous maximum value of the three-phase current in the previous control cycle. The speed tracking error is obtained by comparing the commanded speed with the actual speed.
[0067] When the peak drive current exceeds a preset current threshold or the speed tracking error exceeds a preset error threshold, an anomaly is determined in the parameter transition process, and the current transition is immediately paused. The preset current threshold is 150% of the motor's rated current; in this embodiment, the motor's rated current is 10A, therefore the preset current threshold is 15A. The preset error threshold is 20% of the current commanded speed amplitude.
[0068] When an anomaly occurs, a reverse transition is performed according to the following recursive formula, the mathematical expression of which is: in, The control parameter vector at the moment before transient start-up, and the reverse transient rate coefficient. The reverse transition rate coefficient is set to 0.3 because, in the event of an anomaly, the control parameters need to be rolled back to a safe state in a shorter time to minimize damage to the equipment and workpiece caused by the abnormal operation. The reverse transition also employs an exponential smoothing method to ensure a continuous and smooth rollback process. In this embodiment, the reverse transition completes 90% of the rollback within approximately 3ms.
[0069] S7, Incremental Update of Clustering Model and Online Extension of Policy Library: like Figure 3 As shown, low-confidence samples with a confidence level C ≤ 0.50 are stored in a buffer. The buffer uses a circular buffer structure with a maximum capacity of 500 samples. When the buffer is full, new samples overwrite the oldest samples.
[0070] A batch re-clustering operation is performed every 10,000 control cycles (10 seconds). The batch re-clustering targets all samples in the buffer and the cluster centers of all existing clusters. The samples in the buffer and the existing cluster centers are merged into a single dataset, which is then re-clustered using the standard density peak clustering algorithm. After re-clustering, for each dense region formed, the Euclidean distance from its center to each existing cluster center is calculated. A re-matching distance threshold is set. The value is When the distance from the center of a dense region to the center of an existing cluster is less than When the distance from the center of the dense region to the center of all existing clusters is greater than or equal to the distance from the center of the dense region to the center of all existing clusters, all samples in the dense region are assigned to the existing cluster, and the center and statistic of the cluster are updated using the update method in S2; when the distance from the center of the dense region to the center of all existing clusters is greater than or equal to the distance from the center of the dense region to the center of all existing clusters, all samples in the dense region are assigned to the When this happens, a new working condition category is created.
[0071] After re-clustering is completed, for each newly created cluster, it is determined whether the accumulated number of samples has reached the preset minimum number of training samples. In this embodiment Each sample corresponds to 5 seconds of data. If the number of samples in the new cluster reaches... In the background, training of pre-trained control strategy parameters for this new operating condition category is initiated.
[0072] Background training executes on the controller's idle computing resources without interfering with the real-time operation of the main control loop. Specifically, the main control loop runs as a high-priority task in the real-time operating system, with the highest priority; background training runs as a low-priority task, only consuming CPU resources when the high-priority task is idle. Background training is asynchronous, actively relinquishing CPU resources after each training round and waiting for the next idle window to continue. The training method is the same as the offline pre-training method described in S4 above, using the SoftActor-Critic algorithm in a simulation environment. After training, the cluster center coordinates of the new cluster and the corresponding pre-trained control policy parameters are added to the policy library.
[0073] The continuous evolution of the cluster center adopts an exponentially weighted moving average method, and the update formula is as follows: , where β=0.97. This update is performed each time a sample is assigned to a cluster, allowing the cluster centers to evolve smoothly and continuously in accordance with the long-term, slow changes in operating conditions.
[0074] S8, Strategy Library Maintenance: To prevent the policy library from expanding indefinitely over time, this embodiment also includes a policy library maintenance mechanism.
[0075] A policy library evaluation is performed every 24 hours. The evaluation includes two dimensions: the usage frequency of each policy and the average tracking error during the period in which each policy is invoked. Usage frequency is calculated as the percentage of times each policy is invoked as a target control parameter out of the total number of invocations over the past 24 hours. Average tracking error is calculated as the average root mean square (RMS) value of the tracking error during each invocation of that policy over the past 24 hours.
[0076] Policies used less than 1% of the time and whose average tracking error is more than 1.5 times worse than the global average tracking error are marked as pending elimination. The global average tracking error is the overall average of the root mean square values of tracking errors during the period when all policies were invoked in the past 24 hours.
[0077] Policies marked as pending elimination enter an observation period, which lasts for seven consecutive evaluation cycles (seven consecutive days). If the usage frequency or average tracking error of the policy improves during the observation period, and it no longer meets the elimination criteria, the pending elimination mark is removed. If the policy remains pending elimination for seven consecutive evaluation cycles, it is permanently deleted from the policy library, along with the corresponding operating condition category and all related data.
[0078] In this embodiment, the implementation data for three typical application scenarios are as follows: In Scenario 1, a six-joint industrial robot performs a material handling operation: the robot alternates between three working conditions: unloaded forward transport, heavy-load transport, and unloaded return transport. The data above comes from actual measurements of a six-joint industrial robot prototype in a laboratory environment. Each working condition was repeated 10 times, and the average of the 10 runs was taken as the final result. Before adopting the method of this invention, the speed loop proportional coefficient abruptly changed from 35 for unloaded to 52 for heavy-loaded transport during working condition switching, the current surge peak reached 210% of the rated current, and the end-effector vibration amplitude was 0.15 mm. After adopting the method of this invention, the clustering module completes the identification of the new working condition within approximately 30 ms after the working condition switch, with a confidence level of 0.92. The control parameters smoothly transition to the target value within approximately 6 ms, the current surge peak drops to 135% of the rated current, and the end-effector vibration amplitude drops to 0.06 mm.
[0079] In Scenario 2, a six-joint industrial robot performs high-speed welding: the welding trajectory includes multiple high-speed inflection points, with the speed changing from 0.2 m / s to 1.8 m / s within 0.5 seconds. The above data comes from actual measurements of a standard welding trajectory executed by the same six-joint industrial robot prototype in a laboratory environment, repeated 10 times and the average value taken. Using traditional fixed parameter control, the maximum tracking error at the inflection points is 2.3 mm. After adopting the method of this invention, the online clustering module identifies the transitional conditions during acceleration and deceleration with a confidence level of 0.78. The strategy invocation module adds real-time error fine-tuning to the pre-trained parameters, reducing the maximum tracking error at the inflection points to 0.7 mm.
[0080] In scenario 3, during the collaborative robot assembly operation, the load gradually increases from 0.5 kg to 3.0 kg, and the drive system continuously adjusts the control parameters as the load increases. The above data comes from actual measurements of the collaborative robot prototype in a laboratory environment, with the load increased in 0.5 kg increments, and each load increment repeated 5 times with the average value taken. Using the method of this invention, the clustering model continuously updates the cluster centers during load changes, maintaining a confidence level between 0.7 and 0.9. The control parameters adjust smoothly with load changes, and no parameter oscillations or current surges occur throughout the assembly process. The position tracking accuracy remains stable within ±0.02 mm.
[0081] The implementation data from the above three scenarios show that the AI-adaptive robot equipment precision drive control optimization method provided by this invention can effectively achieve active pre-adaptation and smooth transition of control parameters, significantly improve tracking accuracy and reduce current surges in scenarios with rapid switching of working conditions.
[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. An AI-adaptive precision drive control optimization method for robot equipment, characterized in that, The method includes the following steps: S1, collects the angular velocity, angular acceleration, load torque, d-axis current, q-axis current and current harmonic amplitude of the robot joints in real time, and constructs the working condition feature vector at the current moment; S2, based on the operating condition feature vector, uses the online density peak clustering algorithm to cluster the current operating conditions, and obtains the operating condition category to which the current operating condition belongs and the Mahalanobis distance of the current operating condition feature vector relative to the center of the operating condition category; S3, calculate the confidence level of the current working condition identification based on the Mahalanobis distance and the mean and standard deviation of the Mahalanobis distance of the historical samples of the corresponding working condition category; S4: Pre-trained control strategy parameters corresponding to each working condition category are stored in advance. When the confidence level is greater than the first threshold, the pre-trained control strategy parameter corresponding to the current working condition category is called as the target control parameter. When the confidence level is greater than the second threshold and less than or equal to the first threshold, the pre-trained control strategy parameter corresponding to the current working condition category is called and corrected by combining real-time tracking error as the target control parameter. When the confidence level is less than or equal to the second threshold, the current control parameter remains unchanged. S5, exponentially smooth the transition from the current control parameter to the target control parameter according to the following recursive formula: in, This is the control parameter vector at the current moment. This is the control parameter vector for the next control cycle. Let λ be the target control parameter vector, λ be the preset transient rate coefficient, C be the confidence level, and Δt be the control period.
2. The AI-adaptive robot equipment precision drive control optimization method according to claim 1, characterized in that, The control parameter vector includes the velocity loop proportional coefficient, velocity loop integral coefficient, velocity loop differential coefficient, position loop proportional coefficient, position loop integral coefficient, position loop differential coefficient, velocity feedforward coefficient, acceleration feedforward coefficient, notch filter center frequency, and notch filter depth. The online density peak clustering algorithm uses a time decay factor and a sliding window mechanism for online updates. The operating condition feature vector is an 8-dimensional feature vector, consisting of angular velocity, angular acceleration, load torque, q-axis current, d-axis current, second harmonic amplitude, fourth harmonic amplitude, and sixth harmonic amplitude. The feature values of each dimension are standardized by Z-score and then used in clustering calculations.
3. The AI-adaptive robot equipment precision drive control optimization method according to claim 1, characterized in that, The confidence level is calculated according to the following formula: Where C is the confidence level, and its value ranges from (0,1]. The Mahalanobis distance from the current working condition feature vector to the center of its corresponding working condition category is given. The mean Mahalanobis distance of the historical samples for the corresponding working condition category is updated online using an exponentially weighted moving average. The Mahalanobis distance standard deviation of the historical samples for the corresponding working condition category is updated online using an exponentially weighted moving average.
4. The AI-adaptive robot equipment precision drive control optimization method according to claim 1, characterized in that, The pre-trained control strategy parameters are obtained in the following way: For each work condition category, a flexible motion-evaluation reinforcement learning algorithm is used for offline training in the robot dynamics simulation environment. The reward function is a weighted negative value of the squared tracking error, the squared rate of change of control quantity, and energy consumption. After training convergence, a 9-dimensional control parameter vector is output as the pre-trained control strategy parameters corresponding to that work condition category. The 9-dimensional control parameter vector includes the velocity loop proportional coefficient, velocity loop integral coefficient, velocity loop differential coefficient, position loop proportional coefficient, position loop integral coefficient, position loop differential coefficient, velocity feedforward coefficient, acceleration feedforward coefficient, and notch filter center frequency and notch filter depth.
5. The AI-adaptive robot equipment precision drive control optimization method according to claim 1, characterized in that, The first threshold is 0.85, and the second threshold is 0.50; When the confidence level is greater than 0.85, the pre-trained control policy parameters are directly used as the target control parameters. When the confidence level is greater than 0.50 and less than or equal to 0.85, the pre-trained control policy parameters are corrected according to the following formula and then used as the target control parameters: in, This is the corrected target control parameter vector. Let G be the pre-trained control strategy parameter vector, γ be the preset fine-tuning coefficient, e(t) be the tracking error at the current time, and G be the preset error-parameter gain matrix.
6. The AI-adaptive robot equipment precision drive control optimization method according to claim 1, characterized in that, In step S5, transition safety monitoring is also performed: during the transition of control parameters, the peak value of the drive current and the speed tracking error are monitored in real time; when the peak value of the drive current exceeds a preset current threshold or the speed tracking error exceeds a preset error threshold, the current transition is paused and a reverse transition is performed according to the following recursive formula: in, Control parameter vector at the moment before transition start This is the preset reverse transition rate coefficient.
7. The AI-adaptive robot equipment precision drive control optimization method according to claim 1, characterized in that, S2 also includes the creation of a new working condition category: if the weighted Euclidean distance from the current working condition feature vector to the center of all existing working condition categories is greater than the preset new cluster creation distance threshold, a new working condition category is created with the current working condition feature vector as the initial center. If the number of accumulated samples for a new operating condition category reaches the preset minimum number of training samples, the training of the pre-trained control strategy parameters for that new operating condition category will be started in the background. After the training is completed, the new operating condition category and its corresponding pre-trained control strategy parameters will be added to the pre-stored strategy library.
8. The AI-adaptive robot equipment precision drive control optimization method according to claim 1, characterized in that, The method further includes S6: For each preset evaluation cycle, the usage frequency of the pre-trained control strategy parameters corresponding to each working condition category and the average tracking error during the call period are statistically analyzed. Pre-trained control strategy parameters that are used less frequently than the first frequency threshold and whose average tracking error is worse than a preset multiple of the global average tracking error are marked as to be eliminated. After being in a pending elimination state for several consecutive evaluation cycles, the operating condition category and its corresponding pre-trained control strategy parameters are deleted from the pre-stored strategy library.
9. The AI-adaptive robot equipment precision drive control optimization method according to claim 1, characterized in that, S2 also includes merging and deleting operating condition categories: When the distance between the centers of two existing working condition categories is less than the preset merging distance threshold, the two working condition categories are merged into one working condition category, and the center of the merged category is the weighted average of the centers of the two working condition categories. If no new samples are added to any work condition category within a preset time period, delete that work condition category and its corresponding pre-trained control strategy parameters.
10. The AI-adaptive robot equipment precision drive control optimization method according to claim 1, characterized in that, The method further includes S7: Store the feature vectors of operating conditions with a confidence level less than or equal to the second threshold into the cache area; Each time a preset batch update cycle is run, batch re-clustering is performed on the working condition feature vectors in the cache area, and the newly formed dense region after re-clustering is re-matched with the existing working condition categories or a new working condition category is created. The cluster centers in the online density peak clustering algorithm are updated using an exponentially weighted moving average method. in, This is the updated cluster center vector. t is the cluster center vector before the update, β is the preset cluster center update decay coefficient, and F(t) is the working condition feature vector at the current time.
Citation Information
Patent Citations
Parallel robot self-adaptive adjustment method and system based on attention mechanism
CN120447346A
Intelligent optimization method for control parameters of servo motor system of mobile robot
CN121806499A
