A robot servo control motion feedback optimization system and method

Abstract

The present application relates to the technical field of robot control, and in particular to a robot servo control motion feedback optimization system and method, through collecting joint position, speed, torque and end pose and other multi-source feedback data, a cross-domain motion feedback fusion model is constructed, and spatiotemporal correlation features are extracted in a high-dimensional feature space; a variable structure predictor is used, the prediction time domain length and internal topological structure are adaptively adjusted according to the motion mode, a feedforward compensation signal is generated; then the servo drive instruction is output through dynamic weighted fusion of the feedback signal by a gated fusion network; and the model parameters are continuously optimized by using a double-buffer online incremental learning mechanism. The trajectory tracking accuracy and the calculation resource efficiency are simultaneously optimized in the whole motion mode.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to a robot servo control motion feedback optimization system and method. Background Technology

[0002] Existing robot servo control motion feedback optimization systems typically consist of three core components: a servo driver, a servo motor, and an encoder. The servo motor, acting as an actuator, is responsible for converting control signals into mechanical motion. The servo driver receives instructions and drives the motor. The encoder detects the motor's position and speed in real time and feeds them back to the controller, thereby achieving precise closed-loop control of position, speed, and torque. This three-loop control system, through the cascaded structure of current loop, speed loop, and position loop, enables robots to achieve high-precision motion control under various working conditions. This facilitates the widespread application of servo control systems, powerfully promoting the transformation and upgrading of the manufacturing industry from automation to intelligence, and significantly improving the efficiency and product quality of industrial production. In existing servo control technologies, commonly used motion feedback methods include single feedback forms such as joint position feedback, velocity feedback, and torque feedback. Feedforward-feedback composite control often employs fixed-coefficient addition or mode switching fusion based on simple rules. In predictive control, mainstream methods use predictors with fixed structures, whose prediction time domain length and internal network topology are preset and cannot be dynamically adjusted. They lack adaptive adjustment capabilities under different motion modes such as uniform speed, acceleration / deceleration, start-stop transitions, and complex trajectory tracking. Parameter tuning methods mostly rely on offline calibration or theoretical calculations based on dynamic models, lacking online continuous learning and optimization capabilities. However, in practical industrial applications, robot servo systems face various time-varying conditions such as dynamic load changes, strong coupling of multiple degrees of freedom, nonlinear friction disturbances, and complex trajectory tracking. Existing methods struggle to balance trajectory tracking accuracy and computational resource efficiency across all motion modes—fixed-structure predictors cannot dynamically adapt to the prediction requirements of different conditions, and single feedback modes or simply spliced ​​multi-source data cannot fully exploit the cross-domain spatiotemporal correlation characteristics between motion feedbacks. This severely restricts the trajectory tracking accuracy and robustness of servo control systems in complex dynamic environments. Summary of the Invention

[0003] The purpose of this invention is to provide a method for optimizing motion feedback in robot servo control, comprising the following steps: Collect multi-source motion feedback data during the robot's servo motion process. The multi-source motion feedback data includes at least joint position feedback data, joint velocity feedback data, joint torque feedback data, and end-effector pose feedback data. A cross-domain motion feedback fusion model is constructed, and the multi-source motion feedback data is mapped to a unified high-dimensional feature space for spatiotemporal correlation feature extraction to generate a fusion feedback feature vector. Based on the fused feedback feature vector, an adaptive prediction optimization mechanism is used to generate a feedforward compensation signal. The adaptive prediction optimization mechanism includes a variable structure predictor, which adaptively adjusts its prediction time domain length and internal topology according to the time-varying characteristics of the robot's motion state. The feedforward compensation signal and the feedback control signal are dynamically weighted and fused to generate servo drive control commands. The servo drive control command is output to the robot joint servo driver to drive the robot to perform motion, and the motion execution error is monitored in real time. The motion execution error is fed back to the adaptive prediction optimization mechanism to update the predictor parameters. The actual operating data during the robot's servo motion process is collected, and the internal parameters of the cross-domain motion feedback fusion model and the adaptive prediction optimization mechanism are updated online based on the actual operating data.

[0004] The online incremental update employs a dual-buffer update strategy based on credibility allocation. This involves setting up a first buffer and a second buffer. The first buffer stores the currently collected online running data and constructs an initial update gradient, while the second buffer stores high-value training samples that have undergone quality screening. When the accumulated update gradient in the first buffer reaches a preset threshold, the update gradient is submitted to the second buffer for secondary optimization, and then the model parameters are updated uniformly. The selection of high-value training samples is based on a comprehensive evaluation of sample error magnitude, sample novelty, and sample information entropy.

[0005] The cross-domain motion feedback fusion model includes a spatiotemporal feature encoder and a cross-domain attention fusion module. The spatiotemporal feature encoder is used to independently encode motion feedback data from different domains and generate spatiotemporal feature tensors corresponding to each domain. The cross-domain attention fusion module is used to calculate the correlation weights between feature tensors of different domains through a multi-head attention mechanism, and to perform weighted fusion of feature tensors of each domain based on the correlation weights to generate the fusion feedback feature vector. The dimension of the self-attention matrix of the cross-domain attention fusion module is dynamically adjustable according to the number of input data sources.

[0006] The variable structure predictor includes a motion pattern recognition unit, a prediction time-domain adaptive unit, and a topology selection unit. The motion pattern recognition unit is used to identify the current motion pattern of the robot based on the fused feedback feature vector. The motion patterns include uniform motion mode, acceleration and deceleration motion mode, start-stop transition mode and complex trajectory tracking mode. The prediction time-domain adaptive unit is used to adaptively determine the optimal prediction time-domain length based on the identified motion pattern and the statistical characteristics of the motion execution error. The topology selection unit is used to select the corresponding predictor internal network topology based on the prediction time domain length, wherein a shorter prediction time domain length corresponds to a shallow network structure and a longer prediction time domain length corresponds to a deep network structure.

[0007] The optimal prediction time domain length is determined by solving a multi-objective optimization problem. The optimization objectives of this problem include the cost of prediction and tracking errors, the cost of computational resource consumption, and system stability constraints. The adaptive weight coefficients used in the optimization problem are dynamically adjusted according to the robot's load state and real-time motion complexity.

[0008] The dynamic weighted fusion is implemented through a gated fusion network. The gated fusion network takes the fusion feedback feature vector as input and outputs the fusion weight coefficients of the feedforward compensation signal and the feedback control signal. The fusion weight coefficients satisfy the constraint that the sum is 1 and all are between 0 and 1. In the system startup phase and transient transition phase, the feedforward compensation signal is assigned a higher weight, and in the steady-state operation phase, the feedback control signal is assigned a higher weight.

[0009] The joint torque feedback data is obtained by integrating high-precision torque sensors into each joint of the robot to collect the actual output torque of each joint in real time; at the same time, auxiliary torque data is estimated based on motor current signals; the measured torque of the sensors and the estimated torque of the motor current are fused and verified to generate calibrated joint torque feedback data.

[0010] The end-effector pose feedback data includes visual feedback data and inertial measurement data. The visual feedback data is acquired by a visual sensor installed at the end of the robot or at a fixed external location, and the inertial measurement data is acquired by an inertial measurement unit integrated into the end of the robot. The cross-domain motion feedback fusion model performs spatiotemporal synchronization and joint encoding on the visual feedback data and inertial measurement data to eliminate data mismatch problems caused by different sampling frequencies and data delays.

[0011] The variable structure predictor employs a gradual parameter transition strategy when switching internal topology structures. It uses linear interpolation to gradually transition the current network parameters to the target network parameters to avoid control output jumps caused by topology abrupt changes. The cross-domain motion feedback fusion model also includes a time synchronization unit, which performs timestamp alignment and interpolation processing on multi-source feedback data with different sampling frequencies, and dynamically compensates for the inherent delay of the sensor through an adaptive delay compensation module.

[0012] One of them is a robot servo control motion feedback optimization system, used to implement the robot servo control motion feedback optimization method, characterized in that it includes a multi-source data acquisition module, a cross-domain fusion module, an adaptive prediction module, a dynamic fusion module, a drive execution module, an error monitoring feedback module, and an online learning optimization module; The multi-source data acquisition module has its input terminals connected to the robot's joint position sensor, joint velocity sensor, joint torque sensor, and end-effector pose sensing unit, respectively, for acquiring multi-source motion feedback data. The cross-domain fusion module has its input end connected to the output end of the multi-source data acquisition module, and is used to construct a cross-domain motion feedback fusion model, map the multi-source motion feedback data to a high-dimensional feature space and generate a fusion feedback feature vector. The adaptive prediction module has its first input terminal connected to the output terminal of the cross-domain fusion module and its second input terminal connected to the output terminal of the error monitoring feedback module. It is used to generate a feedforward compensation signal based on the fusion feedback feature vector and motion execution error using an adaptive prediction optimization mechanism. The adaptive prediction optimization mechanism includes a variable structure predictor. The dynamic fusion module has a first input terminal connected to the output terminal of the adaptive prediction module, a second input terminal connected to the feedback control signal source, and a third input terminal connected to the output terminal of the cross-domain fusion module. It is used to dynamically weight and fuse the feedforward compensation signal and the feedback control signal to generate servo drive control commands. The drive execution module has its input end connected to the output end of the dynamic fusion module and its output end connected to the robot joint servo driver, and is used to output the servo drive control command to drive the robot to perform motion. The error monitoring and feedback module has its input end connected to the output end of the drive execution module or the actual motion output end of the robot joint, and is used to monitor the motion execution error in real time, and feed the motion execution error back to the second input end of the adaptive prediction module and the first input end of the online learning and optimization module, respectively. The online learning optimization module has a first input terminal connected to the output terminal of the error monitoring feedback module, a second input terminal connected to the output terminal of the cross-domain fusion module, and a third input terminal connected to the output terminal of the adaptive prediction module. Its output terminal is connected to the parameter update terminal of the cross-domain fusion module and the parameter update terminal of the adaptive prediction module, respectively, for performing online incremental updates of the internal parameters of the cross-domain fusion module and the adaptive prediction module. The online learning optimization module has a first buffer and a second buffer. The first buffer stores the currently collected online running data and constructs an initial update gradient. The second buffer stores high-value training samples that have undergone quality screening. When the update gradient accumulated in the first buffer reaches a preset threshold, the update gradient is submitted to the second buffer for secondary optimization, and then the model parameters are updated uniformly. The selection of high-value training samples is based on a comprehensive evaluation of sample error magnitude, sample novelty, and sample information entropy.

[0013] This invention discloses a robot servo control motion feedback optimization system and method. Taking a six-axis industrial robot in a high-speed gluing or precision assembly scenario as an example, after the system starts, it first collects joint position, velocity, torque, and end-effector pose data in real time through high-precision encoders, torque sensors, and vision and inertial measurement units integrated at the end effector, all installed on the robot's joints. These data come from different physical domains and sampling frequencies; for example, the position data update frequency is 1kHz, and the vision data is 30Hz. Subsequently, a cross-domain motion feedback fusion model maps these heterogeneous data into a unified high-dimensional feature space. It uses a spatiotemporal attention mechanism to mine the potential correlations between different feedback signals, such as the temporal coupling relationship between joint torque fluctuations and end-effector pose deviations, generating a highly condensed fusion feedback feature vector. Based on this, a variable structure predictor dynamically adjusts its prediction time domain length and network depth according to the currently identified motion mode (such as high-speed linear interpolation or circular transition): a shorter prediction time domain and shallower network are used in the rapid start-stop phase to reduce latency, while a longer prediction time domain and deeper network are used in the uniform and stable phase to improve accuracy, thereby generating accurate feedforward compensation signals. The feedforward compensation signal and the traditional PID feedback control signal are dynamically weighted and fused through a gating network. During startup and transient states, the feedforward is the primary method for rapid response, while during steady state, the feedback is the primary method to ensure zero steady-state error tracking. Finally, the command is output to the joint servo driver to drive the motor. At the same time, the error between the actual trajectory and the command trajectory is monitored in real time and immediately fed back to the variable structure predictor to update the predictor's internal state and parameters. In addition, the system continuously collects all data (including input, output, intermediate features and errors) throughout the entire operation process and continuously optimizes the parameters of the cross-domain fusion model and the variable structure predictor through online incremental learning. This makes the robot more and more accurate with continuous operation and can automatically adapt to load changes, joint wear and environmental disturbances. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0015] Figure 1This is a flowchart of the robot servo control motion feedback optimization method of the present invention.

[0016] Figure 2 This is a flowchart of the robot servo control motion feedback optimization method of the present invention.

[0017] Figure 3 This is a flowchart illustrating the internal workflow of the cross-domain motion feedback fusion model of the present invention.

[0018] Figure 4 This is a flowchart of the adaptive variable structure predictor of the present invention.

[0019] Figure 5 This is a flowchart of the online incremental learning double-buffered update process of the present invention.

[0020] Figure 6 This is a schematic diagram of the robot servo control motion feedback optimization system of the present invention.

[0021] In the diagram: 1-Multi-source data acquisition module, 2-Cross-domain fusion module, 3-Adaptive prediction module, 4-Dynamic fusion module, 5-Driven execution module, 6-Error monitoring and feedback module, 7-Online learning and optimization module. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0023] In the description of this invention, it should be understood that "a plurality of" means two or more, unless otherwise explicitly specified.

[0024] Please see Figures 1 to 5 This invention provides a method for optimizing motion feedback in robot servo control, comprising the following steps: S1: Collect multi-source motion feedback data during the robot's servo motion process. The multi-source motion feedback data includes at least joint position feedback data, joint velocity feedback data, joint torque feedback data, and end-effector pose feedback data. S2: Construct a cross-domain motion feedback fusion model, map the multi-source motion feedback data to a unified high-dimensional feature space for spatiotemporal correlation feature extraction, and generate a fusion feedback feature vector; S3: Based on the fused feedback feature vector, an adaptive prediction optimization mechanism is used to generate a feedforward compensation signal. The adaptive prediction optimization mechanism includes a variable structure predictor, which adaptively adjusts its prediction time domain length and internal topology according to the time-varying characteristics of the robot's motion state. S4: Dynamically weight and fuse the feedforward compensation signal and the feedback control signal to generate servo drive control commands; S5: Output the servo drive control command to the robot joint servo driver to drive the robot to perform motion, and monitor the motion execution error in real time. Feed the motion execution error back to the adaptive prediction optimization mechanism to update the predictor parameters. S6: Collect actual running data during the robot's servo motion process, and perform online incremental updates to the internal parameters of the cross-domain motion feedback fusion model and the adaptive prediction optimization mechanism based on the actual running data.

[0025] In this embodiment, taking a six-axis industrial robot in a high-speed gluing or precision assembly scenario as an example, after the system is started, it first collects joint position, speed, torque and end pose data in real time through high-precision encoders, torque sensors and vision and inertial measurement units integrated at the end of the robot installed in each joint. These data come from different physical domains and sampling frequencies, for example, the position data update frequency is 1kHz and the vision data is 30Hz.

[0026] Subsequently, the cross-domain motion feedback fusion model maps these heterogeneous data into a unified high-dimensional feature space and uses a spatiotemporal attention mechanism to mine the potential correlations between different feedback signals, such as the temporal coupling relationship between joint torque fluctuations and end-effector pose deviations, generating a highly condensed fusion feedback feature vector.

[0027] Based on this, the variable structure predictor dynamically adjusts its prediction time domain length and network depth according to the currently identified motion mode (such as high-speed linear interpolation or circular transition): a shorter prediction time domain and shallower network are used during rapid start-up and shutdown phases to reduce latency, while a longer prediction time domain and deeper network are used during steady-state phases to improve accuracy, thereby generating a precise feedforward compensation signal. This feedforward compensation signal is dynamically weighted and fused with the traditional PID feedback control signal through a gating network. During startup and transient states, feedforward is the primary method for achieving rapid response, while feedback is the primary method for ensuring zero steady-state error tracking during steady-state states.

[0028] The final command is output to the joint servo driver to drive the motor. At the same time, the error between the actual trajectory and the command trajectory is monitored in real time, and the error is immediately fed back to the variable structure predictor to update the predictor's internal state and parameters.

[0029] In addition, the system continuously collects all data (including inputs, outputs, intermediate features and errors) throughout the entire operation process, and continuously optimizes the parameters of the cross-domain fusion model and variable structure predictor through online incremental learning, so that the robot becomes more and more accurate with continuous operation and can automatically adapt to load changes, joint wear and environmental disturbances.

[0030] Furthermore, the online incremental update adopts a dual-buffer update strategy based on credibility allocation, namely, setting a first buffer and a second buffer. The first buffer is used to store the currently collected online running data and construct the initial update gradient, while the second buffer is used to store high-value training samples that have undergone quality screening. When the update gradient accumulated in the first buffer reaches a preset threshold, the update gradient is submitted to the second buffer for secondary optimization, and then the model parameters are updated uniformly. The selection of high-value training samples is based on a comprehensive evaluation of sample error magnitude, sample novelty, and sample information entropy.

[0031] In this embodiment, to avoid model performance degradation caused by gradient noise and sample redundancy during online learning, the system adopts a double-buffered update strategy. The first buffer temporarily stores the most recent 1000 running data in a first-in, first-out manner. Each time a new sample is received, the system calculates an initial update gradient based on the current model.

[0032] When the accumulated gradients in the first buffer (e.g., the average of the gradient norms) exceed a preset threshold, the system does not immediately update the model. Instead, it submits these gradients to the second buffer. The second buffer stores high-value training samples that have undergone quality screening: the system evaluates the error magnitude (whether the trajectory tracking error is abnormally large), novelty (the degree of difference between the current feature and the existing sample distribution), and information entropy (the uncertainty of the model's prediction for that sample) of each sample. Only samples with the highest combined scores of these three metrics are retained.

[0033] Subsequently, the system uses carefully selected samples from the second buffer to perform secondary optimization of the gradient, such as through momentum averaging or adaptive learning rate adjustment, before finally performing a stable update on the parameters of the cross-domain fusion model and the variable structure predictor. This mechanism effectively prevents abnormal transient data from damaging the model, while ensuring that the model can quickly absorb representative new knowledge.

[0034] Furthermore, the cross-domain motion feedback fusion model includes a spatiotemporal feature encoder and a cross-domain attention fusion module. The spatiotemporal feature encoder is used to independently encode motion feedback data from different domains and generate spatiotemporal feature tensors corresponding to each domain. The cross-domain attention fusion module is used to calculate the correlation weights between feature tensors of different domains through a multi-head attention mechanism, and to perform weighted fusion of feature tensors of each domain based on the correlation weights to generate the fusion feedback feature vector. The dimension of the self-attention matrix of the cross-domain attention fusion module is dynamically adjustable according to the number of input data sources.

[0035] In this embodiment, the spatiotemporal feature encoder designs a dedicated encoding network for each type of feedback data: for one-dimensional time-series signals such as joint position and velocity, a 1D convolutional neural network is used to extract local time-series patterns; for joint torque, the current estimation value and the sensor measurement value are fused and encoded; for end pose, visual data is extracted for spatial features through 2D convolution, and inertial data is extracted for motion features through a fully connected network.

[0036] The encoded feature tensors for each domain are aligned in both the temporal and spatial dimensions. Subsequently, the cross-domain attention fusion module uses these tensors as input to a multi-head attention mechanism, calculating the correlation weights between each pair of domains—for example, during high-speed motion, the attention weight between joint velocity features and end-effector inertia features increases significantly, reflecting the strong coupling between them. The dimension of the self-attention matrix is ​​dynamically adjusted based on the number of data sources actually connected. When the visual sensor temporarily fails due to lighting issues, the module automatically reduces the dimension, fusing only the remaining data sources to ensure system robustness. The fused feedback feature vector output after weighted fusion retains the key information of each domain and their spatiotemporal correlations, providing rich state representations for subsequent predictions.

[0037] Furthermore, the variable structure predictor includes a motion pattern recognition unit, a prediction time-domain adaptive unit, and a topology selection unit. The motion pattern recognition unit is used to identify the current motion pattern of the robot based on the fused feedback feature vector. The motion patterns include uniform motion mode, acceleration and deceleration motion mode, start-stop transition mode and complex trajectory tracking mode. The prediction time-domain adaptive unit is used to adaptively determine the optimal prediction time-domain length based on the identified motion pattern and the statistical characteristics of the motion execution error. The topology selection unit is used to select the corresponding predictor internal network topology based on the prediction time domain length, wherein a shorter prediction time domain length corresponds to a shallow network structure and a longer prediction time domain length corresponds to a deep network structure.

[0038] In this embodiment, the motion pattern recognition unit uses a lightweight classification network, taking a sliding window that fuses feedback feature vectors as input, and outputs the probability of the current motion pattern in real time. For example, when the robot starts to accelerate from a standstill, the velocity value in the feature vector increases rapidly from zero, and the torque spikes. The classifier judges it as a "start-stop transition mode". When the speed is stable and the error variance is extremely small, it is judged as a "uniform speed motion mode".

[0039] The predictive time-domain adaptive unit determines the optimal prediction time domain length based on the identified pattern and the root mean square value of the motion execution error in the recent period through fuzzy logic rules: the start-stop transition mode uses a time domain length of 5~10 milliseconds, the acceleration-deceleration mode uses 10~20 milliseconds, the constant speed mode uses 20~50 milliseconds, and the complex trajectory tracking mode is dynamically adjusted according to the trajectory curvature.

[0040] The topology selection unit selects a suitable internal topology for the variable structure predictor from a pre-stored pool of network structures based on a determined time-domain length: a shallow gated recurrent unit network with only 1-2 hidden layers is used for short time-domain scenarios, while a deep Transformer structure with residual connections is used for long time-domain scenarios. This adaptive mechanism ensures that the system achieves the best balance between prediction accuracy and computational latency under different operating conditions.

[0041] Furthermore, the optimal prediction time domain length is determined by solving a multi-objective optimization problem. The optimization objectives of this problem include the cost of prediction and tracking errors, the cost of computational resource consumption, and system stability constraints. The adaptive weight coefficients used in the optimization problem are dynamically adjusted according to the robot's load state and real-time motion complexity.

[0042] In this embodiment, the system applied to the method does not use a fixed rule table to retrieve the prediction time domain. Instead, it solves a constrained multi-objective optimization problem in real time. Specifically, for each candidate time domain length, the system estimates the prediction tracking error for several future steps when using that time domain (by quickly evaluating it using a lightweight meta-model). At the same time, it measures the CPU / GPU computation time and memory usage required to perform the prediction for that length and evaluates the stability margin of the closed-loop system in that time domain (e.g., based on the Lyapunov exponent).

[0043] The optimization objective is to minimize the weighted sum of prediction tracking error and computational resource consumption while ensuring stability constraints are met. The adaptive weighting coefficients are not fixed: when the robot's load suddenly increases (e.g., when it grasps a heavy object), the system automatically increases the weight of prediction tracking error, as accurate prediction becomes more critical; when the robot's motion complexity is low (e.g., simple reciprocating motion), the system increases the weight of computational resource consumption, prioritizing reduced energy consumption and computational power usage. Through this dynamic multi-objective optimization, the variable structure predictor can find the most suitable prediction time domain length for the current working condition within milliseconds.

[0044] Furthermore, the dynamic weighted fusion is implemented through a gated fusion network. The gated fusion network takes the fusion feedback feature vector as input and outputs the fusion weight coefficients of the feedforward compensation signal and the feedback control signal. The fusion weight coefficients satisfy the constraint that the sum is 1 and all are between 0 and 1. In the system startup phase and transient transition phase, the feedforward compensation signal is assigned a higher weight, and in the steady-state operation phase, the feedback control signal is assigned a higher weight.

[0045] In this embodiment, a lightweight gated fusion network is introduced. This network takes the fused feedback feature vector at the current moment as input, passes it through a two-neuron fully connected network with a softmax output layer, and directly generates feedforward weights α and feedback weights β, where α + β = 1. When the robot has just started or when the motion command changes abruptly, the velocity and acceleration terms in the fused feature vector are relatively large. The gated network automatically sets α to above 0.8, making the control command mainly based on feedforward compensation to achieve rapid following. When the robot enters the constant speed or precise positioning stage, the error integral term in the fused feature vector becomes dominant. The gated network increases β to above 0.9, making feedback control play a major role and eliminating steady-state error. This dynamic fusion method allows the system to improve dynamic response by utilizing the predictive nature of feedforward while retaining the robustness of feedback to suppress disturbances, which is significantly better than fixed-coefficient composite control.

[0046] Furthermore, the joint torque feedback data is obtained in the following way: high-precision torque sensors are integrated into each joint of the robot to collect the actual output torque of each joint in real time; at the same time, auxiliary torque data is estimated based on motor current signals; the measured torque of the sensors and the estimated torque of the motor current are fused and verified to generate calibrated joint torque feedback data.

[0047] In this embodiment, to obtain accurate and reliable joint torque information, the system employs a dual redundancy and fusion verification strategy. Each robot joint's output end is equipped with a strain gauge torque sensor to directly measure the actual output torque. This data has high bandwidth and good accuracy, but is susceptible to temperature drift.

[0048] Meanwhile, the servo driver's current loop outputs the motor phase current in real time, and an auxiliary torque value is estimated by the motor torque constant and friction model. This estimated value is not affected by temperature but will accumulate deviation with model error.

[0049] The system feeds these two torque data streams into a Kalman filter for fusion: in steady state, the actual sensor measurements are prioritized; when abnormal spikes appear in the sensor signal, the estimated values ​​are referenced for smoothing; and the zero drift and gain error of the sensor are estimated online, outputting calibrated torque feedback data. This fusion method ensures the accuracy of the torque feedback while improving its reliability and anti-interference capability, providing high-quality input for subsequent dynamic feedforward compensation.

[0050] Furthermore, the end-effector pose feedback data includes visual feedback data and inertial measurement data. The visual feedback data is acquired through a visual sensor installed at the end of the robot or at a fixed external location, and the inertial measurement data is acquired through an inertial measurement unit integrated into the end of the robot. The cross-domain motion feedback fusion model performs spatiotemporal synchronization and joint encoding on the visual feedback data and inertial measurement data to eliminate data mismatch problems caused by different sampling frequencies and data delays.

[0051] In this embodiment, the precise pose of the robot's end effector is crucial for precision assembly and trajectory tracking. This solution integrates two complementary sensing methods: vision and inertial. Vision sensors (such as binocular cameras or structured light cameras) can provide absolute pose information, but the update frequency is low (30~60Hz) and there is a processing delay; the inertial measurement unit (IMU) provides three-axis acceleration and angular velocity, with an update frequency as high as 1kHz, but there is integral drift.

[0052] The system first synchronizes the two types of data to the same time base through timestamp alignment and interpolation. Then, it uses an adaptive delay compensation module to estimate the delay time from acquisition to availability of visual data and performs forward prediction on inertial data to compensate for this delay.

[0053] Subsequently, the spatiotemporal feature encoder jointly encodes the synchronized visual features (such as keypoint coordinates) and inertial features (such as pose quaternions) to generate a high-frequency, low-latency end-effector pose estimation sequence, which serves as part of the fused feedback feature vector. This fusion enables the robot to maintain precise control over the end-effector even when its vision is obstructed or it is moving rapidly, relying on inertia.

[0054] Furthermore, when switching internal topology, the variable structure predictor adopts a parameter gradual transition strategy, using linear interpolation to gradually transition the current network parameters to the target network parameters, in order to avoid control output jumps caused by topology abrupt changes. The cross-domain motion feedback fusion model also includes a time synchronization unit, which performs timestamp alignment and interpolation processing on multi-source feedback data with different sampling frequencies, and dynamically compensates for the inherent delay of the sensor through an adaptive delay compensation module.

[0055] In this embodiment, when the variable structure predictor needs to switch from a shallow network to a deep network (e.g., from start-stop mode to uniform long-time domain prediction mode), directly replacing the network weights will cause a step change in the prediction output, which in turn will cause a jump in the servo control command and make the robot jitter.

[0056] To this end, the system adopts a parameter gradual transition strategy: in the first N control cycles after the switch (N is usually set to 10~50), the new network parameters gradually transition from the old parameters to the target parameters through linear interpolation, that is, the current parameter = (1-λ)×old parameter + λ×target parameter, and λ increases linearly from 0 to 1.

[0057] This strategy smooths out changes in the predictor's output, avoiding control shocks. Simultaneously, the time synchronization unit in the cross-domain motion feedback fusion model manages the time of all input data: it maintains a timestamp queue for each data source and uses an asynchronous interpolation algorithm to unify all data onto the same time grid. The adaptive delay compensation module identifies the delay time of each sensor channel in real time based on historical data (e.g., 30ms delay for vision processing, 2ms delay for torque sensor filtering) and uses a Smith predictor or forward predictor to dynamically compensate for the delay, ensuring that the fused feature vector truly reflects the robot's current state, rather than its past state. These two design details significantly improve the stability and real-time performance of the entire control system.

[0058] Please see Figure 6 A robot servo control motion feedback optimization system is used to implement the robot servo control motion feedback optimization method, including a multi-source data acquisition module 1, a cross-domain fusion module 2, an adaptive prediction module 3, a dynamic fusion module 4, a drive execution module 5, an error monitoring feedback module 6, and an online learning optimization module 7. The multi-source data acquisition module 1 has its input terminals connected to the robot's joint position sensor, joint velocity sensor, joint torque sensor, and end-effector pose sensing unit, respectively, for acquiring multi-source motion feedback data. The cross-domain fusion module 2 has its input end connected to the output end of the multi-source data acquisition module 1, and is used to construct a cross-domain motion feedback fusion model, map the multi-source motion feedback data to a high-dimensional feature space and generate a fusion feedback feature vector. The adaptive prediction module 3 has its first input terminal connected to the output terminal of the cross-domain fusion module 2 and its second input terminal connected to the output terminal of the error monitoring feedback module 6. It is used to generate a feedforward compensation signal based on the fusion feedback feature vector and motion execution error using an adaptive prediction optimization mechanism. The adaptive prediction optimization mechanism includes a variable structure predictor. The dynamic fusion module 4 has its first input terminal connected to the output terminal of the adaptive prediction module 3, its second input terminal connected to the feedback control signal source, and its third input terminal connected to the output terminal of the cross-domain fusion module 2. It is used to dynamically weight and fuse the feedforward compensation signal and the feedback control signal to generate servo drive control commands. The drive execution module 5 has its input end connected to the output end of the dynamic fusion module 4, and its output end connected to the robot joint servo driver, for outputting the servo drive control command to drive the robot to perform motion. The error monitoring feedback module 6 has its input end connected to the output end of the drive execution module 5 or the actual motion output end of the robot joint, and is used to monitor the motion execution error in real time, and feed the motion execution error back to the second input end of the adaptive prediction module 3 and the first input end of the online learning optimization module 7 respectively. The online learning optimization module 7 has its first input terminal connected to the output terminal of the error monitoring feedback module 6, its second input terminal connected to the output terminal of the cross-domain fusion module 2, and its third input terminal connected to the output terminal of the adaptive prediction module 3. Its output terminal is connected to the parameter update terminal of the cross-domain fusion module 2 and the parameter update terminal of the adaptive prediction module 3, respectively, for performing online incremental updates of the internal parameters of the cross-domain fusion module 2 and the adaptive prediction module 3. The online learning optimization module 7 is internally configured with a first buffer and a second buffer. The first buffer is used to store the currently collected online running data and construct a preliminary update gradient. The second buffer is used to store high-value training samples that have undergone quality screening. When the update gradient accumulated in the first buffer reaches a preset threshold, the update gradient is submitted to the second buffer for secondary optimization, and then the model parameters are updated uniformly. The selection of high-value training samples is based on a comprehensive evaluation of sample error amplitude, sample novelty, and sample information entropy.

[0059] In this embodiment, the above modules are closely connected according to the data flow and control flow to form a complete closed-loop optimization system. The multi-source data acquisition module 1 acquires raw data from hardware sensors in real time, performs preliminary filtering and format unification on it, and then sends it to the cross-domain fusion module 2.

[0060] The cross-domain fusion module 2 completes spatiotemporal alignment, feature encoding, and attention fusion, and outputs a high-dimensional fusion feedback feature vector.

[0061] The feature vector is simultaneously fed into the adaptive prediction module 3 and the dynamic fusion module 4: the adaptive prediction module 3 uses the fusion feature and the motion execution error returned from the error monitoring feedback module 6 to calculate the feedforward compensation signal through the variable structure predictor; the dynamic fusion module 4 then dynamically determines the weights of feedforward and feedback through a gating network based on the fusion feature to synthesize the final servo drive control command.

[0062] The drive execution module 5 converts the commands into PWM or analog signals and outputs them to the servo driver to drive the robot's movement. The error monitoring and feedback module 6 collects the deviation between the actual position / velocity and the command in real time. On the one hand, it directly feeds this deviation back to the adaptive prediction module 3 for predictor state updates; on the other hand, it sends this deviation, along with other operating data, to the online learning and optimization module 7.

[0063] The online learning optimization module 7 uses a double buffering mechanism to perform quality screening and gradient accumulation on the data stream, and periodically updates the deep learning model parameters in the cross-domain fusion module 2 and the adaptive prediction module 3.

[0064] The entire system completes a full-link cycle of "perception-fusion-prediction-decision-execution-feedback-learning" in each control cycle, enabling the robot servo control to continuously evolve and adapt to complex and ever-changing industrial environments, significantly improving trajectory tracking accuracy, dynamic response speed, and system robustness.

[0065] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for optimizing motion feedback in robot servo control, characterized in that, Includes the following steps: Collect multi-source motion feedback data during the robot's servo motion process. The multi-source motion feedback data includes at least joint position feedback data, joint velocity feedback data, joint torque feedback data, and end-effector pose feedback data. A cross-domain motion feedback fusion model is constructed, and the multi-source motion feedback data is mapped to a unified high-dimensional feature space for spatiotemporal correlation feature extraction to generate a fusion feedback feature vector. Based on the fused feedback feature vector, an adaptive prediction optimization mechanism is used to generate a feedforward compensation signal. The adaptive prediction optimization mechanism includes a variable structure predictor, which adaptively adjusts its prediction time domain length and internal topology according to the time-varying characteristics of the robot's motion state. The feedforward compensation signal and the feedback control signal are dynamically weighted and fused to generate servo drive control commands. The servo drive control command is output to the robot joint servo driver to drive the robot to perform motion, and the motion execution error is monitored in real time. The motion execution error is fed back to the adaptive prediction optimization mechanism to update the predictor parameters. The actual operating data during the robot's servo motion process is collected, and the internal parameters of the cross-domain motion feedback fusion model and the adaptive prediction optimization mechanism are updated online based on the actual operating data.

2. The robot servo control motion feedback optimization method as described in claim 1, characterized in that, The online incremental update adopts a dual-buffer update strategy based on credibility allocation, namely, setting up a first buffer and a second buffer. The first buffer is used to store the currently collected online running data and construct the initial update gradient, while the second buffer is used to store high-value training samples that have undergone quality screening. When the update gradient accumulated in the first buffer reaches a preset threshold, the update gradient is submitted to the second buffer for secondary optimization, and then the model parameters are updated uniformly. The selection of high-value training samples is based on a comprehensive evaluation of sample error magnitude, sample novelty, and sample information entropy.

3. The robot servo control motion feedback optimization method as described in claim 1, characterized in that, The cross-domain motion feedback fusion model includes a spatiotemporal feature encoder and a cross-domain attention fusion module. The spatiotemporal feature encoder is used to independently encode motion feedback data from different domains and generate spatiotemporal feature tensors corresponding to each domain. The cross-domain attention fusion module is used to calculate the correlation weights between feature tensors of different domains through a multi-head attention mechanism, and to perform weighted fusion of feature tensors of each domain based on the correlation weights to generate the fusion feedback feature vector. The dimension of the self-attention matrix of the cross-domain attention fusion module is dynamically adjustable according to the number of input data sources.

4. The robot servo control motion feedback optimization method as described in claim 1, characterized in that, The variable structure predictor includes a motion pattern recognition unit, a prediction time-domain adaptive unit, and a topology selection unit. The motion pattern recognition unit is used to identify the current motion pattern of the robot based on the fused feedback feature vector. The motion patterns include uniform motion mode, acceleration and deceleration motion mode, start-stop transition mode and complex trajectory tracking mode. The prediction time-domain adaptive unit is used to adaptively determine the optimal prediction time-domain length based on the identified motion pattern and the statistical characteristics of the motion execution error. The topology selection unit is used to select the corresponding predictor internal network topology based on the prediction time domain length, wherein a shorter prediction time domain length corresponds to a shallow network structure and a longer prediction time domain length corresponds to a deep network structure.

5. The robot servo control motion feedback optimization method as described in claim 4, characterized in that, The optimal prediction time domain length is determined by solving a multi-objective optimization problem. The optimization objectives of this problem include the cost of prediction and tracking errors, the cost of computational resource consumption, and system stability constraints. The adaptive weight coefficients used in the optimization problem are dynamically adjusted according to the robot's load state and real-time motion complexity.

6. The robot servo control motion feedback optimization method as described in claim 1, characterized in that, The dynamic weighted fusion is implemented through a gated fusion network. The gated fusion network takes the fusion feedback feature vector as input and outputs the fusion weight coefficients of the feedforward compensation signal and the feedback control signal. The fusion weight coefficients satisfy the constraint that the sum is 1 and all are between 0 and 1. In the system startup phase and transient transition phase, the feedforward compensation signal is assigned a higher weight, and in the steady-state operation phase, the feedback control signal is assigned a higher weight.

7. The robot servo control motion feedback optimization method as described in claim 1, characterized in that, The joint torque feedback data is obtained in the following way: high-precision torque sensors are integrated into each joint of the robot to collect the actual output torque of each joint in real time; at the same time, auxiliary torque data is estimated based on motor current signals; the measured torque of the sensors and the estimated torque of the motor current are fused and verified to generate calibrated joint torque feedback data.

8. The robot servo control motion feedback optimization method as described in claim 1, characterized in that, The end-effector pose feedback data includes visual feedback data and inertial measurement data. The visual feedback data is acquired by a visual sensor installed at the end of the robot or at a fixed external location, and the inertial measurement data is acquired by an inertial measurement unit integrated into the end of the robot. The cross-domain motion feedback fusion model performs spatiotemporal synchronization and joint encoding on the visual feedback data and inertial measurement data to eliminate data mismatch problems caused by different sampling frequencies and data delays.

9. The robot servo control motion feedback optimization method as described in claim 1, characterized in that, When switching internal topology, the variable structure predictor adopts a parameter gradual transition strategy, which uses linear interpolation to gradually transition the current network parameters to the target network parameters in order to avoid control output jumps caused by topology abrupt changes. The cross-domain motion feedback fusion model also includes a time synchronization unit, which performs timestamp alignment and interpolation processing on multi-source feedback data with different sampling frequencies, and dynamically compensates for the inherent delay of the sensor through an adaptive delay compensation module.

10. A robot servo control motion feedback optimization system, used to implement the robot servo control motion feedback optimization method as described in claim 1, characterized in that, It includes a multi-source data acquisition module, a cross-domain fusion module, an adaptive prediction module, a dynamic fusion module, a drive execution module, an error monitoring and feedback module, and an online learning and optimization module; The multi-source data acquisition module has its input terminals connected to the robot's joint position sensor, joint velocity sensor, joint torque sensor, and end-effector pose sensing unit, respectively, for acquiring multi-source motion feedback data. The cross-domain fusion module has its input end connected to the output end of the multi-source data acquisition module, and is used to construct a cross-domain motion feedback fusion model, map the multi-source motion feedback data to a high-dimensional feature space and generate a fusion feedback feature vector. The adaptive prediction module has its first input terminal connected to the output terminal of the cross-domain fusion module and its second input terminal connected to the output terminal of the error monitoring feedback module. It is used to generate a feedforward compensation signal based on the fusion feedback feature vector and motion execution error using an adaptive prediction optimization mechanism. The adaptive prediction optimization mechanism includes a variable structure predictor. The dynamic fusion module has a first input terminal connected to the output terminal of the adaptive prediction module, a second input terminal connected to the feedback control signal source, and a third input terminal connected to the output terminal of the cross-domain fusion module. It is used to dynamically weight and fuse the feedforward compensation signal and the feedback control signal to generate servo drive control commands. The drive execution module has its input end connected to the output end of the dynamic fusion module and its output end connected to the robot joint servo driver, and is used to output the servo drive control command to drive the robot to perform motion. The error monitoring and feedback module has its input end connected to the output end of the drive execution module or the actual motion output end of the robot joint, and is used to monitor the motion execution error in real time, and feed the motion execution error back to the second input end of the adaptive prediction module and the first input end of the online learning and optimization module, respectively. The online learning optimization module has a first input terminal connected to the output terminal of the error monitoring feedback module, a second input terminal connected to the output terminal of the cross-domain fusion module, and a third input terminal connected to the output terminal of the adaptive prediction module. Its output terminal is connected to the parameter update terminal of the cross-domain fusion module and the parameter update terminal of the adaptive prediction module, respectively, for performing online incremental updates of the internal parameters of the cross-domain fusion module and the adaptive prediction module. The online learning optimization module is internally configured with a first buffer and a second buffer. The first buffer is used to store the currently collected online running data and construct the initial update gradient. The second buffer is used to store high-value training samples that have undergone quality screening. When the update gradient accumulated in the first buffer reaches a preset threshold, the update gradient is submitted to the second buffer for secondary optimization, and then the model parameters are updated uniformly. The selection of high-value training samples is based on a comprehensive evaluation of sample error magnitude, sample novelty, and sample information entropy.

Images