Generalization generation method for industrial robot pose trajectory supporting variable operation path points

Abstract

The present invention relates to the field of robot trajectory planning, and in particular to a generalization generation method for an industrial robot pose trajectory supporting variable operation path points. The method comprises: collecting multiple sets of robot end pose trajectories; aligning time steps of the pose trajectories by means of a multi-dimensional dynamic time warping algorithm; constructing a robot pose trajectory Gaussian mixture model by combining a variational Bayesian method and a pose quaternion tangent space mapping method; calculating a reference pose trajectory distribution by means of Gaussian mixture regression; performing kernelized representation on the pose trajectory distribution, and solving for an optimal hyperparameter of a kernel function by minimizing a root-mean-square error of a reproduced reference trajectory; and by updating the reference pose trajectory distribution, generating a robot end pose trajectory adapted to operation path points. In the present invention, generalization generation of an industrial robot pose trajectory supporting variable operation path points is achieved, thereby improving the degree of flexibility of an operation path of an industrial robot.

Description

Generalized generation method of pose trajectory for industrial robots supporting variable operation path points

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on September 6, 2023, with application number 202311149325.1 and invention name “Generalized Generation Method of Industrial Robot Posture Trajectory Supporting Variable Operation Path Points”, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present invention relates to the field of robot trajectory planning, and in particular to a generalized generation method for an industrial robot posture trajectory supporting variable operation path points. Background Art

[0003] Since their initial appearance, industrial robots have become an integral part of manufacturing, automating repetitive, dangerous, and heavy-duty tasks, improving production efficiency while reducing costs and risks. Currently, most industrial robots installed on production floors are limited to specific tasks and lack intelligence. However, with the increasing demand for product diversification, the flexibility of industrial robots' operations needs to increase to accommodate increasingly flexible production environments.

[0004] Providing robots with easy-to-program interfaces is a prerequisite for their widespread use in manufacturing. However, programming difficulty is a limiting factor. Existing teach-in programming or offline programming methods require professional engineers or developers to do a lot of work for a given task. While these methods can ensure high-precision completion of tasks within a single, fixed operating scenario, they suffer from long deployment times, low programming efficiency, and high operator professionalism requirements. When the robot's operating scenario changes, it becomes unable to perform the task and requires re-teaching. Consequently, it cannot adapt to the changing operating scenarios of flexible manufacturing systems, hindering the further application of robots.

[0005] Demonstration learning provides a method for improving a robot's adaptability to different operating scenarios. It aims to allow non-expert users to teach robots how to perform different tasks through demonstration. Imitation learning of robot motion trajectories is a key research direction in demonstration learning. This approach can learn motion features from a small number of teaching trajectories. By generalizing these motion features to new operating scenarios, the robot can adapt to dynamic changes in the operating scenario, thereby achieving flexibility in the robot's operating path.

[0006] Currently, robot trajectory learning technologies are primarily categorized into dynamic motion primitives, probabilistic motion primitives, and kernelized motion primitives. These methods enable robots to adapt to changing pathpoints, but they rarely consider the generation of robot posture trajectories. Existing trajectory learning methods often project posture quaternions onto a single Euclidean tangent space and then use Gaussian mixture models for probabilistic modeling. This ignores the inherent geometric constraints of posture quaternions, making it difficult to generate high-precision posture trajectories. Therefore, current learning-based trajectory generation technologies still have many shortcomings. On the one hand, the use of traditional Gaussian mixture models fails to consider the geometric constraints of the robot's posture trajectory, resulting in low probabilistic modeling accuracy. On the other hand, they cannot simultaneously generate both the robot's position trajectory and posture trajectory.

[0007] Summary of the Invention

[0008] The present invention proposes a method for generating an industrial robot pose trajectory that supports variable operation path points. This method can overcome the shortcomings of existing teaching programming methods. Based on the probabilistic modeling of the robot pose trajectory, the reference pose trajectory distribution is calculated through Gaussian mixture regression. The pose trajectory adapted to the operation path points is generated through the kernelized representation of the pose trajectory distribution, thereby improving the flexibility of the robot's operation path.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] S1: Collect multiple sets of robot end pose trajectories;

[0011] S2: Align the time steps of multiple sets of robot end pose trajectories to obtain the robot end pose trajectory alignment set;

[0012] S3: Based on the robot end pose trajectory alignment set, a Gaussian mixture model of the robot pose trajectory is constructed by combining variational Bayesian and pose quaternion tangent space mapping methods;

[0013] S4: Based on the robot pose trajectory Gaussian mixture model, the reference pose trajectory distribution is obtained by combining the pose quaternion covariance parallel migration method in tangent space and Gaussian mixture regression calculation;

[0014] S5: Optimally kernelize the reference pose trajectory distribution to obtain the optimal hyperparameters of the kernel function;

[0015] S6: Update the reference pose trajectory distribution according to the robot's operation path points, and generate the robot's end pose trajectory adapted to the operation path points after kernelizing the current reference pose trajectory distribution according to the optimal hyperparameters of the kernel function.

[0016] In S2, a multi-dimensional dynamic time warping method is used to set the distance between trajectory points to ensure the synchronization of position and posture dimensions and to satisfy the posture quaternion geometric constraints, thereby aligning the time steps of multiple groups of robot terminal posture trajectories.

[0017] The S3 is specifically:

[0018] Based on the robot end position trajectory alignment set, a Gaussian mixture model of the robot position trajectory is constructed using the variational Bayesian method;

[0019] Based on the robot end pose trajectory alignment set, the Gaussian mixture model of the robot pose trajectory is constructed using the pose quaternion tangent space mapping method;

[0020] The robot posture trajectory Gaussian mixture model is composed of the robot position trajectory Gaussian mixture model and the robot posture trajectory Gaussian mixture model.

[0021] The S4 is specifically:

[0022] Calculate the robot end reference position trajectory based on Gaussian mixture regression method;

[0023] The covariance of each Gaussian component in the Gaussian mixture model of the robot posture trajectory is parallel migrated on the tangent space of the posture quaternion, and the covariance ∑ ||i Replace the covariance calculation of each Gaussian component in the traditional Gaussian mixture regression to obtain the covariance component of the i-th Gaussian component posture dimension After traversing and calculating all Gaussian components, the covariance matrix consisting of the covariance components corresponding to all Gaussian components in the posture dimension is obtained. According to the covariance matrix Calculate the mean matrix of the posture dimension to obtain the end reference posture trajectory;

[0024] The reference pose trajectory distribution is composed of the end reference position trajectory and the end reference pose trajectory.

[0025] In S5, the optimal hyperparameters of the kernel function are obtained by minimizing the root mean square error of the distribution of the reproduced reference pose trajectory.

[0026] The variational Bayesian method in S3 aims to automatically optimize the number of Gaussian components of the pose trajectory Gaussian mixture model, use the variational Bayesian method to learn a whole class of probability models, calculate the posterior probability on the position trajectory data for each probability model, and infer the parameter π in the Gaussian mixture model through the variational Bayesian method. k 、μ k and Σ k The maximum posterior distribution of .

[0027] The attitude quaternion tangent space mapping method in S3 refers to projecting the attitude quaternion from its Riemann space to the Euclidean tangent space of another attitude quaternion, and can restore it from the tangent space back to the Riemann space through the tangent space inverse mapping.

[0028] The method of parallel migration of attitude quaternion covariance in tangent space in S4 refers to migrating the covariance of attitude quaternions in the tangent space of two attitude quaternions, ensuring that the dispersion of different attitude quaternions is correctly represented in the tangent space.

[0029] The pose trajectory distribution kernel representation method in S5 refers to kernel representation of the pose trajectory distribution by using a kernel method and minimizing the KL divergence between the reference pose trajectory distribution and the parameterized trajectory distribution;

[0030] For the kernelized representation of the posture trajectory distribution, first the reference posture trajectory Projected to the starting point q of the posture trajectory s In order to ensure the consistency of the attitude quaternion covariance in different tangent spaces, the covariance of the reference attitude trajectory is From the mean The tangent space is parallel to the starting point q of the posture trajectory s The cutting space.

[0031] The root mean square error of the reproduced reference trajectory in S5 refers to penalizing the difference between the reproduced trajectory and the reference trajectory by squaring the Euclidean distance in the position dimension and the rotation distance in the attitude dimension.

[0032] The reference pose trajectory distribution update in S6 refers to updating the reference pose trajectory distribution RD using the work path point VD, and finding the reference pose trajectory distribution RD for each work path point. The covariance of each operation path point in VD is a diagonal matrix, and the diagonal element is a very small constant, which indicates the accuracy of the trajectory to adapt to the operation path point. The accuracy of the position dimension and the attitude dimension can be set separately; represents the Euclidean distance between two working path points, t r Represents the corresponding moment of the reference trajectory point, and Δ represents a very small constant.

[0033] The beneficial effects of the present invention are:

[0034] The present invention combines the variational Bayesian method with the attitude quaternion tangent space mapping method to construct a Gaussian mixture model of the robot posture trajectory, which improves the accuracy of the posture trajectory probability modeling while considering the geometric constraints of the robot posture trajectory.

[0035] The present invention takes into account the generalized generation of position trajectories and posture trajectories at the same time. The generated trajectories can adapt to changing operation path points, and the operation path of the robot is more flexible.

[0036] In summary, the present invention achieves the generalized generation of an industrial robot posture trajectory that supports variable operation path points, thereby improving the flexibility of the industrial robot's operation path. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] FIG1 is a schematic flow diagram of the method of the present invention.

[0038] FIG2 is a flowchart of the algorithm for constructing a Gaussian mixture model of the robot posture trajectory in step S3.

[0039] FIG3 is a schematic diagram of the reference position trajectory calculated in step S4 .

[0040] FIG4 is a flow chart of the algorithm for updating the reference pose trajectory distribution in step S62.

[0041] FIG5 is a schematic diagram of the end position trajectory of the industrial robot generated in step S63. DETAILED DESCRIPTION

[0042] The present invention will be further described below with reference to the accompanying drawings and specific examples.

[0043] As shown in FIG1 , the present invention uses the barrel handling process in nuclear waste treatment by an industrial robot as an example for explanation, and specifically includes the following steps:

[0044] S1: Collect multiple sets of robot end-position trajectories through drag teaching, offline programming, virtual teaching, etc.

[0045] The robot is, for example, an industrial robot, such as an ABB IRB6700 industrial robot, but is not limited thereto.

[0046] Drag teaching in S1 refers to the demonstration of the user completing a specific task by physically moving the robot's joints in gravity compensation mode, while recording the position, posture and joint angle of the end effector through its own sensors, or recording the robot's surrounding environment information (such as obstacle positions, the status of other collaborative robots or users) through the external vision system.

[0047] Offline programming in S1 refers to creating a virtual robot operation scene in offline programming software such as ABB's RobotStudio and simulating and verifying the planned trajectory.

[0048] Virtual teaching in S1 refers to the convenient and intuitive control of a virtual environment or a remote robot through virtual reality, augmented reality, etc. to complete the collection of the robot's terminal posture trajectory.

[0049] S2: Align the time steps of multiple sets of robot end pose trajectories to obtain the robot end pose trajectory alignment set.

[0050] In S2, the time length of the robot end pose trajectory collected each time is different, which makes it inconvenient to subsequently model the probability of the pose trajectory. Therefore, a multi-dimensional dynamic time warping method is used to set the distance between trajectory points to ensure the synchronization of position and pose dimensions and satisfy the pose quaternion geometric constraints, thereby aligning the time steps of multiple groups of robot end pose trajectories.

[0051] S2 is specifically:

[0052] S21: Convert the robot's end posture into a unit quaternion, which is a Riemannian manifold on a four-dimensional sphere.

[0053] S22: Calculate the distance between custom pose trajectory points. The distance between two trajectory points in position and pose dimensions is the Euclidean distance and the rotation distance respectively. The rotation distance of the two pose quaternions is the L2 regularization||log of the tangent space projection. f (g)||2. Where g and f represent a posture quaternion respectively.

[0054] S23: Use dynamic programming algorithm to calculate the distance matrix consisting of the distance between each pair of data points.

[0055] S24: Solve the alignment path between two trajectory sequences according to the distance matrix. L =(N,M), calculate p in reverse order l-1 , get the optimal alignment path p * , satisfying p * =(p1,…,p L ), where p L Map coordinates for trajectory points, N and M are the time series lengths of the two trajectory sequences, and L represents the length of the optimal alignment path.

[0056] S25: Align all pose trajectories with the target pose trajectory to obtain multiple sets of robot end pose trajectories after time alignment.

[0057] S3: Based on the robot end pose trajectory alignment set, a Gaussian mixture model of the robot pose trajectory is constructed by combining the variational Bayesian and pose quaternion tangent space mapping methods.

[0058] As shown in Figure 2, S3 is specifically:

[0059] S31: Based on the robot end pose trajectory alignment set, a Gaussian mixture model of the robot position trajectory is constructed using the variational Bayes method.

[0060] S32: Based on the robot end pose trajectory alignment set, the Gaussian mixture model of the robot pose trajectory is constructed using the pose quaternion tangent space mapping method.

[0061] S321: Pass responsibility coefficient r nk Weighted pose trajectory point x nq Projection in the mean tangent space to update the Gaussian component mean μ kq .

[0062] S322: Gaussian component mean μ kq The dispersion of the attitude quaternion is estimated as the covariance ∑ of the attitude quaternion in the tangent space kq .

[0063] The robot posture trajectory Gaussian mixture model is composed of the robot position trajectory Gaussian mixture model and the robot posture trajectory Gaussian mixture model.

[0064] S4: Based on the robot pose trajectory Gaussian mixture model, combined with the attitude quaternion covariance parallel migration method in tangent space and Gaussian mixture regression calculation, the reference pose trajectory distribution is obtained.

[0065] S4 is specifically:

[0066] S41: The end reference position trajectory of the robot bucket handling task is calculated based on the traditional Gaussian mixture regression method, as shown in Figure 3.

[0067] S42: Parallel migrate the covariance of each Gaussian component in the robot posture trajectory Gaussian mixture model on the tangent space of the posture quaternion, and use the covariance ∑ ||i Replace the covariance calculation of each Gaussian component in the traditional Gaussian mixture regression to obtain the covariance component of the i-th Gaussian component posture dimension After traversing and calculating all Gaussian components, the covariance matrix consisting of the covariance components corresponding to all Gaussian components in the posture dimension is obtained. According to the covariance matrix The mean matrix of the posture dimension is calculated to obtain the end reference posture trajectory.

[0068] S43: The reference position trajectory and the reference posture trajectory of the end are used to form a reference posture trajectory distribution. Each trajectory point in the calculated reference posture trajectory distribution is Contains time point t nr , end position mean Terminal position variance End position mean and the terminal position variance

[0069] S5: Perform optimal kernelization on the reference pose trajectory distribution to obtain the optimal hyperparameters of the kernel function.

[0070] S5 is specifically:

[0071] S51: The position trajectory distribution of the terminal is kernelized by the kernel method and minimizing the KL divergence between the reference pose trajectory distribution and the parameterized trajectory distribution.

[0072] S52: For the kernel representation of the terminal's posture trajectory distribution, first convert the terminal's reference posture trajectory Projected to the starting point q of the posture trajectory s The cutting space.

[0073] S53: In order to ensure the consistency of the attitude quaternion covariance in different tangent spaces, the covariance of the end reference attitude trajectory is From the mean The tangent space is parallel to the starting point q of the posture trajectory s The cutting space.

[0074] S54: Select an appropriate kernel function from Gaussian kernel, Cauchy kernel, and periodic kernel for the robot's bucket handling task.

[0075] S55: The difference between the reproduced trajectory and the reference trajectory is penalized by squaring the Euclidean distance in the position dimension and the rotation distance in the attitude dimension and recording them as the root mean square error.

[0076] S56: Find the optimal hyperparameters of the kernel function by minimizing the root mean square error of the reproducing end reference pose trajectory.

[0077] S6: Update the reference pose trajectory distribution according to the robot's operation path points, and generate the robot's end pose trajectory adapted to the operation path points after kernelizing the current reference pose trajectory distribution according to the optimal hyperparameters of the kernel function.

[0078] S6 specifically:

[0079] S61: Set the operation path points according to the changes in the robot bucket handling operation scene. Each operation path point is represented by in, Indicates the expected time to pass the operation path point, and represent the desired end position and posture respectively, and They respectively represent the covariance of the adapted operation path points. The covariance is a diagonal matrix with a very small constant as the diagonal element, which indicates the accuracy of the trajectory adapting to the operation path points. The accuracy of the position dimension and the attitude dimension can be set separately.

[0080] S62: For each desired operation path point, update the reference pose trajectory distribution according to the process in Figure 4.

[0081] S63: Generate a robot end pose trajectory adapted to the work path point using the kernelized representation of the pose trajectory distribution. The generated end pose trajectory of the industrial robot is shown in FIG5 .

[0082] During the barrel handling process of nuclear waste disposal by an industrial robot, the robot's workspace is limited, and the starting position of the barrel on the workbox support platform varies. This requires certain positioning requirements for the barrel on the workbox support platform, while also ensuring that the robot does not collide with the surrounding environment. Therefore, the robot needs to have flexible operating capabilities. The terminal position trajectory generated by the present invention enables the industrial robot to adapt to the changing operating path points during the barrel handling process, improving the flexibility of its operating path.

[0083] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modifications and changes made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A generalized generation method for an industrial robot posture trajectory supporting variable operation path points, characterized in that: The following steps are involved: S1: Collect multiple sets of robot terminal position trajectories; S2: Align the time steps of multiple groups of robot terminal pose trajectories to obtain the robot terminal pose trajectory alignment set; S3: According to the robot terminal pose trajectory alignment set, the robot pose trajectory Gaussian mixture model is constructed by combining variational Bayesian and pose quaternion tangent space mapping methods; S4: Based on the robot posture trajectory Gaussian mixture model, the reference posture trajectory distribution is obtained by combining the parallel migration method of the posture quaternion covariance in the tangent space and the Gaussian mixture regression calculation; S5: Optimally kernelize the reference pose trajectory distribution to obtain the optimal hyperparameters of the kernel function; S6: Update the reference pose trajectory distribution according to the robot's operating path points, and generate the robot's terminal pose trajectory adapted to the operating path points after kernelizing the current reference pose trajectory distribution according to the optimal hyperparameters of the kernel function.

2. According to claim 1, a generalized generation method for an industrial robot posture trajectory supporting variable operation path points is characterized in that: In S2, a multi-dimensional dynamic time warping method is used to set the distance between trajectory points to ensure the synchronization of position and attitude dimensions and to satisfy the geometric constraints of attitude quaternion, thereby aligning the time steps of multiple groups of robot terminal posture trajectories.

3. The generalized generation method of an industrial robot posture trajectory supporting variable operation path points according to claim 1 is characterized in that: The S3 is specifically: Based on the robot end position trajectory alignment set, the Gaussian mixture model of the robot position trajectory is constructed using the variational Bayes method; Based on the alignment set of the robot's terminal posture trajectory, the Gaussian mixture model of the robot's posture trajectory is constructed using the posture quaternion tangent space mapping method; The Gaussian mixture model of the robot's position trajectory and the Gaussian mixture model of the robot's posture trajectory are composed of the Gaussian mixture model of the robot's position trajectory and the Gaussian mixture model of the robot's posture trajectory.

4. The generalized generation method of an industrial robot posture trajectory supporting variable operation path points according to claim 1 is characterized in that: The S4 is specifically: Calculate the robot end reference position trajectory based on Gaussian mixture regression method; The covariance of each Gaussian component in the robot posture trajectory Gaussian mixture model is converted into the posture quaternion Parallel migration on the tangent space, using the covariance Σ on the tangent space after parallel migration ||i Replace the covariance calculation of each Gaussian component in the traditional Gaussian mixture regression to obtain the covariance component of the i-th Gaussian component attitude dimension After traversing and calculating all Gaussian components, the covariance matrix consisting of the covariance components corresponding to all Gaussian components in the posture dimension is obtained. According to the covariance matrix Calculate the mean matrix of the posture dimension to obtain the end reference posture trajectory; The reference posture trajectory distribution is composed of the end reference position trajectory and the end reference posture trajectory.

5. The generalized generation method of an industrial robot posture trajectory supporting variable operation path points according to claim 1, characterized in that: In S5, the optimal hyperparameters of the kernel function are obtained by minimizing the root mean square error of the distribution of the reproduced reference pose trajectory.

Images