An online prediction compensation method for parallel robots based on OnlineSNAIL-MER

Abstract

This invention discloses an online prediction and compensation method for parallel robots based on OnlineSNAIL-MER, comprising: first, controlling the moving platform of a Stewart parallel robot to output continuous fixed-point motion according to a given nominal pose; second, building an online measurement system for the moving platform pose, and solving the moving platform pose online based on a forward kinematics model; third, constructing an online regression prediction model for the pose error of the Stewart parallel robot, OnlineSNAIL-MER, which combines OnlineSNAIL and MER; and finally, pre-compensating the online predicted pose error to the nominal pose online, with the compensated nominal pose serving as the input to the Stewart parallel robot controller. This method solves the problems of dynamic, time-varying, and random errors that offline calibration cannot handle. It is simple to operate, does not require periodic model retraining, and enables the Stewart parallel robot to achieve high-precision, long-term stable motion under non-preset working conditions.

Description

Technical Field

[0001] This invention belongs to the field of robot precision control and online meta-learning technology, specifically involving an online prediction and compensation method for pose error of Stewart parallel robot based on OnlineSNAIL-MER network, which enables the robot to maintain high accuracy when the environment changes without the need for frequent model retraining or manual intervention in the compensation process. Background Technology

[0002] The Stewart parallel robot, consisting of six independent kinematic chains connecting the moving and stationary platforms to form a closed-loop structure, boasts advantages such as high load-bearing capacity, high precision, and fast response speed. It plays a crucial role in applications such as precision machining and assembly, medical robots, and support and adjustment systems for astronomical telescopes. Absolute pose accuracy is a critical performance indicator for the Stewart parallel robot, determining whether a six-DOF parallel robot can be applied to high-precision positioning tasks. In practical applications, six-DOF parallel robots are affected by geometric and non-geometric factors, leading to discrepancies between the actual pose and the nominal pose input by the control system.

[0003] Traditional pose error compensation methods establish a system error model to identify key geometric or motion parameters, thereby calculating the compensation amount and performing offline compensation before system operation. However, they cannot accurately calculate non-geometric errors caused by factors such as joint deformation and component wear. Data-driven deep learning-based compensation methods effectively solve the problem of non-geometric error compensation. By building a neural network to construct a mapping relationship between nominal pose and pose error, the input commands of the control system are corrected in reverse, thus achieving pose error compensation.

[0004] Most existing deep learning compensation methods employ offline models, requiring the construction of datasets and offline training. This results in fixed model parameters and weights, lacking adaptability and hindering lifelong learning. In contrast, online meta-learning methods can not only quickly adapt to individual tasks amidst a continuous stream of new tasks but also continuously optimize their internal learning algorithms, thereby improving their efficiency in adapting to future tasks. This approach provides a novel adaptive and intelligent paradigm for online accuracy compensation of the Stewart parallel robot. Its core lies in continuously optimizing the model online to address the dynamic non-geometric error problem that traditional compensation methods struggle to handle. Summary of the Invention

[0005] The purpose of this invention is to provide an online prediction and compensation method for pose error of Stewart parallel robots based on the OnlineSNAIL-MER network, aiming to solve the problem of difficulty in compensating for dynamic non-geometric errors, so as to enable Stewart parallel robots to maintain high-precision long-term stable operation.

[0006] The technical solution of this invention: A method for online prediction and online compensation of pose error of Stewart parallel robot based on OnlineSNAIL-MER network, which includes the following steps:

[0007] S1: Controls the moving platform of the Stewart parallel robot to output continuous fixed-point motion, and each point of the continuous fixed-point motion is a given nominal pose;

[0008] S2: Build an online pose measurement system for the Stewart parallel robot moving platform, and use the real-time collected extension and retraction of the electric cylinders to solve the actual pose of the moving platform online based on the forward kinematics model;

[0009] S3: Construct an online regression prediction model for the pose error of the Stewart parallel robot, combining OnlineSNAIL and MER. The input of the model is the nominal pose of the Stewart parallel robot motion platform, and the output is the predicted value of the pose error. The data is input and output in a streaming manner.

[0010] S4: Based on the real-time data interaction function between the controller and the model, the online predicted pose error is pre-compensated to the nominal pose online.

[0011] In the aforementioned method for online prediction and compensation of pose error of Stewart parallel robot based on OnlineSNAIL-MER network, in step S1, an inverse kinematics model is constructed based on the closed-loop vector method. According to the given nominal pose, the target length of each electric cylinder is solved inversely. Each electric cylinder is driven to reach the calculated target length through closed-loop servo control, and closed-loop control is maintained after reaching the target length to keep the pose stable.

[0012] In the aforementioned method for online prediction and compensation of pose error of Stewart parallel robot based on OnlineSNAIL-MER network, in step S2, an online pose measurement system is built. This system includes six measurement branches, which act as follower devices and move along with the extension and retraction of the electric cylinder. The system is integrated with the Stewart parallel robot to collect the extension and retraction of the electric cylinder in real time. Based on the forward kinematics model, the actual pose of the platform is solved online.

[0013] In the aforementioned method for online prediction and compensation of pose error of Stewart parallel robot based on OnlineSNAIL-MER network, in step S3, the pose error online regression prediction model OnlineSNAIL-MER is an online meta-learning model. It is a model that trains a sub-model and then updates the main model. The main model and the sub-model have the same structure, which is composed of TCBlock and AttentionBlock modules stacked alternately. The input is a six-dimensional pose data and the output is a six-dimensional pose error data, that is, the online predicted pose error.

[0014] In the aforementioned method for online prediction and compensation of pose error of Stewart parallel robot based on OnlineSNAIL-MER network, in step S4, the pose error predicted online by the model is stored in the variable of the control program through the real-time data interaction function between the controller and the model. The variable changes with the change of the predicted pose error, and the nominal pose is obtained by subtracting the variable from the nominal pose.

[0015] The benefits of this invention are as follows: Compared with the prior art, this invention can significantly improve the absolute pose accuracy of Stewart parallel robots through online measurement, online prediction and online compensation. At the same time, it provides an intelligent adaptive compensation strategy to effectively deal with the influence of dynamic non-geometric errors and meet the requirements of high-precision long-term stable operation.

[0016] This invention has the following advantages:

[0017] (1) The present invention has built an online measurement system that can measure the extension and retraction of the electric cylinder in real time and calculate the actual pose of the Stewart parallel robot online;

[0018] (2) The present invention constructs an OnlineSNAIL-MER online meta-learning model, which can quickly learn the complex mapping relationship between nominal pose and pose error under the condition of few samples, and continuously optimize itself to realize online prediction of pose error;

[0019] (3) The present invention designs an online posture error compensation system for Stewart parallel robot, which integrates online measurement device with online prediction model to realize online pre-compensation of posture error. Attached Figure Description

[0020] Appendix Figure 1 This is a flowchart of the method of the present invention;

[0021] Appendix Figure 2 This is a structural diagram of the OnlineSNAIL-MER online meta-learning model of the method of this invention;

[0022] Appendix Figure 3 This study compares the pose error of the Stewart parallel robot after compensation with the actual pose error under sudden load conditions based on this method. Detailed Implementation

[0023] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. The present application will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] Figure 1 This invention illustrates an online prediction and compensation method for pose error of a Stewart parallel robot based on an OnlineSNAIL-MER network. The method includes the following steps:

[0025] S1: Controls the moving platform of the Stewart parallel robot to output continuous fixed-point motion, and each point of the continuous fixed-point motion is a given nominal pose.

[0026] In this embodiment, the host computer sends drive control commands to the controller, which then sends control signals to the six actuators. The actuators drive the corresponding electric cylinders according to the commands, thereby achieving synchronous extension and retraction of the six legs of the Stewart parallel robot, causing the moving platform to undergo pose changes. To enable the moving platform to complete continuous fixed-point motion according to a given nominal pose sequence, inverse kinematics analysis is used to resolve each pose into the target extension and retraction amount corresponding to the six electric cylinders. The extension and retraction amount is calculated according to the following formula:

[0027]

[0028] Where L is the extension or displacement of the electric cylinder, usually in mm; n is the number of revolutions of the servo motor; and p is the lead of the lead screw, in mm / r. Based on this formula, by writing the corresponding control program, each electric cylinder can be precisely driven to the specified length, ultimately realizing the continuous movement of the moving platform along the predetermined nominal pose.

[0029] S2: Build an online pose measurement system for the Stewart parallel robot moving platform. Based on the forward kinematics model, the system will collect the extension and retraction of the electric cylinders in real time and solve the actual pose of the moving platform online.

[0030] In this embodiment, six metrology branches, each composed of a grating ruler and a reading head, are used as servo devices, connected to the moving and stationary platforms of the Stewart parallel robot via ball joints. The actual pose of the moving platform is acquired using a laser tracker, and 42 geometric parameters of the metrology branches are calculated using the least squares method, thereby calibrating the geometric error of the metrology branches. The measurement accuracy of the calibrated metrology branches is essentially consistent with that of the laser tracker. The data acquisition module of the controller acquires the readings of the metrology branches in real time, i.e., the absolute position of the electric cylinders. Based on the forward kinematics model, the actual pose of the moving platform is solved online.

[0031] S3: Construct an online regression prediction model for the pose error of the Stewart parallel robot, combining OnlineSNAIL and MER. The input of the model is the nominal pose of the Stewart parallel robot motion platform, and the output is the predicted value of the pose error. The data is input and output in a streaming manner.

[0032] In this embodiment, the OnlineSNAIL-MER model structure is as follows: Figure 2 As shown, the input is the nominal pose input in the form of a data stream, and the output is the predicted pose error value corresponding to this nominal pose. Only after the Stewart parallel robot completes a fixed-point motion will the next nominal pose be input into the model. Before the next nominal pose is input into the model, the control system transmits the actual pose data collected in real time by the online measurement system to the model. The model then calculates the actual pose error, which is used as a label to update the model parameters. The OnlineSNAIL model consists of three alternately stacked causal temporal convolutional TCBlocks and a self-attention mechanism AttentionBlock. The MER meta-optimization mechanism is an online continuous learning strategy. The outer loop corresponds to the main model update in the online phase; that is, after each receiving of the current sample, the model first performs a regular gradient descent to adapt to the latest data. The inner loop builds a model replica based on this and uses historical samples from the experience replay pool for multi-step training, allowing the replica model to fully adapt to the previous data distribution. A Reptile-style meta-update is used, and the meta-update formula is:

[0033]

[0034] Where θ represents the parameters of the main model after being updated in the outer loop; β represents the sub-model parameters after the inner loop; β is the meta-learning step size. Based on formula (2), the main model parameters are interpolated and adjusted along the "direction of change of sub-model parameters relative to main model parameters", so as to maintain the learning effect of historical data while continuously learning new data and reduce model performance degradation. The model uses the Adam optimizer, the learning rate is set to 0.0001, and the loss function is mean squared error. Online normalization adopts the Welford method, and the number of warm-up samples is 15. The capacity of the experience replay pool is 15, and no more than 15 historical samples are used for each meta-learning. The sub-model is trained on the replay samples for 50 rounds, and the meta-update step size β is set to 0.3. The evaluation index of the model's prediction pose error is shown in Table 1. The results show that the model has a fast convergence speed (CS), a smoothness (JR) close to 0, that is, the prediction effect is relatively smooth, and the final accuracy (FA) and the overall accuracy (LA) are very high, indicating that the model has excellent prediction performance.

[0035] Table 1 Model Prediction Performance

[0036]

[0037] S4: Based on the real-time data interaction function between the controller and the model, the online predicted pose error is pre-compensated to the nominal pose online.

[0038] In this embodiment, the OnlineSNAIL-MER model synchronously transmits the online predicted pose error to the control system. The control system uses this predicted value as a compensation value to compensate for the given value. Subsequently, the Stewart parallel robot executes the compensated fixed-point motion, and this process is repeated until the Stewart parallel robot completes the continuous fixed-point motion after online compensation. When the Stewart parallel robot encounters a sudden load during operation, this online compensation method can quickly adapt to the new environment. Compensation Results Figure 3 As shown in the figure. The results show that the proposed method can effectively reduce pose error, has adaptive compensation effect, and is suitable for long-term high-precision compensation work of Stewart parallel robots.

[0039] The foregoing description provides a detailed account of embodiments of the present invention and is not intended to limit the invention in any way. Those skilled in the art can make various optimizations, improvements, and modifications based on this invention. Therefore, the scope of protection of this invention should be defined by the appended claims.

Claims

1. An OnlineSNAIL-MER based parallel robot online prediction compensation method, characterized in that, Includes the following steps: S1: Control the moving platform of the Stewart parallel robot to perform continuous fixed-point motion according to the given nominal pose sequence; S2: Build an online posture measurement system for the moving platform, collect the extension and retraction of the electric cylinder in real time, and calculate the actual posture of the moving platform online based on the positive kinematics model; S3: Construct an online meta-learning model based on OnlineSNAIL and MER, OnlineSNAIL-MER. This model consists of a main model and a sub-model. Both have the same structure, consisting of TCBlock and AttentionBlock stacked alternately. The model takes the nominal pose as input and outputs the corresponding six-dimensional pose error prediction value in real time. After each fixed-point motion is completed, the deviation between the actual pose obtained by the online measurement system and the nominal pose is used as a label to perform online meta-updates on the model, specifically including: S3-1: Construct an experience replay pool to store the most recent N sets of historical sample data. Each set of samples includes the nominal pose and its corresponding actual pose error. S3-2: Outer loop update: After receiving the new sample, the main model first performs a regular gradient descent update to adapt to the latest data distribution; S3-3: Inner loop update: Copy the parameters of the main model to build a sub-model, randomly draw historical samples from the experience replay pool, and train the sub-model in multiple rounds to make it fully adapt to the historical data distribution; S3-4: Meta-parameter update: Using a Reptile-style meta-learning strategy, the trained sub-model parameters are interpolated and fused with the current main model parameters to update the main model parameters. The meta-update formula is: ； where θ is the parameter of the main model after the outer loop update; is the parameter of the sub-model after the inner loop update; and β is the meta-learning step size S3-5: Apply the updated main model parameter to the subsequent online pose error prediction. S4: Real-time compensation of the online predicted pose error to the nominal pose, generating the compensated pose command and inputting it to the Stewart parallel robot controller to achieve online pre-compensation of pose error.

2. The OnlineSNAIL-MER based parallel robot online predictive compensation method according to claim 1, characterized in that: In step S1, an inverse kinematics model is constructed based on the closed-loop vector method. The target length of each electric cylinder is solved inversely according to the given nominal pose. Each electric cylinder is driven to reach the calculated target length through closed-loop servo control, and closed-loop control is maintained after reaching the target length to keep the pose stable.

3. The OnlineSNAIL-MER based parallel robot online predictive compensation method according to claim 1, wherein: In step S2, an online pose measurement system is built. This system includes six measurement branches, which act as follower devices and move in accordance with the extension and retraction of the electric cylinder. The system is integrated with the Stewart parallel robot to collect the extension and retraction of the electric cylinder in real time. Based on the forward kinematics model, the actual pose of the platform is solved online.

4. The OnlineSNAIL-MER based parallel robot online predictive compensation method of claim 1, wherein: In step S3, the pose error online regression prediction model OnlineSNAIL-MER is an online meta-learning model. It is a model that trains a sub-model and then updates the main model. The main model and the sub-model have the same structure, which is composed of TCBlock and AttentionBlock modules stacked alternately. The input is a six-dimensional pose data and the output is a six-dimensional pose error data, that is, the pose error predicted online.

5. The OnlineSNAIL-MER based parallel robot online predictive compensation method of claim 1, wherein: In step S4, the pose error predicted online by the model is stored in the variable of the control program through the real-time data interaction function between the controller and the model. The variable changes with the change of the predicted pose error, and the nominal pose is obtained by subtracting the variable from the nominal pose.

6. The online prediction and compensation method for parallel robots based on OnlineSNAIL-MER according to claim 1, characterized in that, In step S1, an inverse kinematics model is constructed based on the closed-loop vector method. The target extension and retraction of each electric cylinder is solved inversely according to the nominal pose. The electric cylinder is driven to reach the target length through closed-loop servo control to maintain pose stability.

7. The online prediction and compensation method for parallel robots based on OnlineSNAIL-MER according to claim 1, characterized in that, In step S2, the online measurement system includes six metering branches, which are installed in parallel with the electric cylinder as servo devices to collect the extension and retraction of the electric cylinder in real time, and calculate the actual pose of the moving platform through a positive kinematics model.

8. The online prediction and compensation method for parallel robots based on OnlineSNAIL-MER according to claim 1, characterized in that, In step S3, the OnlineSNAIL-MER model consists of a main model and a sub-model. Both have the same structure, consisting of TCBlock and AttentionBlock stacked alternately. The model input is a six-dimensional nominal pose, and the output is a six-dimensional pose error prediction value. The model uses a meta-optimization mechanism for online updates, and the parameters of the main model are adjusted by interpolation based on the training results of the sub-model on historical samples.

Images