Supports a generalized generation method for industrial robot pose trajectories with variable work path points.

Abstract

This invention discloses a generalized generation method for industrial robot pose trajectories that supports variable work path points. First, multiple sets of robot end-effector pose trajectories are collected; then, time step alignment of the pose trajectories is achieved using a multidimensional dynamic time warping algorithm; a Gaussian mixture model of the robot pose trajectory is constructed by combining variational Bayesian and attitude quaternion tangent space mapping methods; a reference pose trajectory distribution is calculated using Gaussian mixture regression; the pose trajectory distribution is kernelized, and the optimal hyperparameters of the kernel function are solved by minimizing the root mean square error of the reproduced reference trajectory; finally, an end-effector pose trajectory adapted to the work path points is generated by updating the reference pose trajectory distribution. This invention achieves generalized generation of industrial robot pose trajectories that supports variable work path points, thereby improving the flexibility of industrial robot work paths.

Description

Technical Field

[0001] This invention relates to a trajectory generation method in the field of robot trajectory planning, and more particularly to a generalized generation method for industrial robot pose trajectories that supports variable operation path points. Background Technology

[0002] Since their inception, industrial robots have become an indispensable part of manufacturing, automating repetitive, hazardous, and heavy-duty tasks while improving production efficiency and reducing costs and risks. Currently, most industrial robots installed in production workshops can only perform specific tasks and have a low level of intelligence. However, with the increasing demand for product diversification, the flexibility of industrial robot operations needs to be improved to adapt to increasingly flexible production environments.

[0003] Providing robots with easy-to-program interfaces is a prerequisite for their widespread application in manufacturing. However, programming difficulty is a limiting factor. Existing teach-in programming or offline programming requires specialized engineers or developers to do a lot of work on a given task. While it can ensure high-precision completion of tasks in a single, fixed work environment, it suffers from problems such as long deployment time, low programming efficiency, and high requirements for operator skills. When the robot's work environment changes, it cannot perform the task and needs to be re-taught. Therefore, it cannot adapt to the changing work environments in flexible manufacturing systems, all of which hinder the further application of robots.

[0004] Demonstration learning provides a method to improve the adaptability of robots to different work scenarios, aiming to allow non-professional users to teach robots the execution process of different tasks through demonstrations. Imitation learning of robot motion trajectories is an important research direction of demonstration learning. It can learn motion features from a small number of taught trajectories, and by generalizing the motion features in new work scenarios, the robot can adapt to the dynamic changes in the work scenario, thereby achieving flexibility in the robot's work path.

[0005] Currently, robot trajectory learning techniques mainly fall into categories such as dynamic motion primitives, probabilistic motion primitives, and kernelized motion primitives. These methods enable robots to adapt to changing work path points, 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, ignoring the inherent geometric constraints of posture quaternions, making it difficult to generate high-precision posture trajectories. Therefore, current learning-based trajectory generation techniques still have many shortcomings. On the one hand, the use of traditional Gaussian mixture models does not consider the geometric constraints of robot pose trajectories, and the accuracy of probabilistic modeling is low. On the other hand, they cannot consider the simultaneous generation of robot position and posture trajectories. Summary of the Invention

[0006] To address the problems existing in current industrial robot trajectory planning technology based on teach-in programming, this invention proposes a method for generating industrial robot pose trajectories that supports variable work path points.

[0007] The method of this invention can overcome the shortcomings of existing teaching programming methods. Based on the probabilistic modeling of robot pose trajectory, it calculates the reference pose trajectory distribution through Gaussian mixture regression, and generates pose trajectories adapted to the operation path points through the kernel representation of the pose trajectory distribution, thereby improving the flexibility of the robot's operation path.

[0008] The technical solution adopted in this invention is:

[0009] S1: Collect multiple sets of robot end-effector pose trajectories;

[0010] S2: Align the time steps of multiple sets of robot end-effector pose trajectories to obtain the robot end-effector pose trajectory alignment set;

[0011] S3: Based on the robot end-effector pose trajectory alignment set, a Gaussian mixture model of robot pose trajectory is constructed by combining variational Bayesian and pose quaternion tangent space mapping methods.

[0012] S4: Based on the Gaussian mixture model of robot pose trajectory, combined with the method of parallel transfer of attitude quaternion covariance in tangent space and Gaussian mixture regression calculation, the reference pose trajectory distribution is obtained;

[0013] S5: Perform optimal kernel representation on the reference pose trajectory distribution to obtain the optimal hyperparameters of the kernel function;

[0014] S6: Update the reference pose trajectory distribution based on the robot's work path points, and generate the robot end pose trajectory adapted to the work path points by kernelizing the current reference pose trajectory distribution according to the optimal hyperparameters of the kernel function.

[0015] In S2, the multidimensional dynamic time warping method is used to ensure the synchronization of position and attitude dimensions and to satisfy the attitude quaternion geometric constraints by setting the distance between trajectory points, thereby aligning the time steps of multiple sets of robot end-effector pose trajectories.

[0016] Specifically, S3 is:

[0017] S31: Based on the robot end-effector pose trajectory alignment set, a Gaussian mixture model of the robot position trajectory is constructed using the variational Bayesian method;

[0018] S32: Based on the robot end-effector pose trajectory alignment set, a Gaussian mixture model of robot pose trajectory is constructed using the pose quaternion tangent space mapping method.

[0019] The robot pose trajectory Gaussian mixture model is composed of the robot position trajectory Gaussian mixture model and the robot attitude trajectory Gaussian mixture model.

[0020] Specifically, S4 is:

[0021] S41: Calculate the reference position trajectory of the robot end effector based on the Gaussian mixture regression method;

[0022] S42: Parallel transfer the covariance of each Gaussian component in the Gaussian mixture model of the robot pose trajectory over the tangent space of the pose quaternions, and use the covariance in the parallel transferred tangent space. The covariance component of the pose dimension of the i-th Gaussian component is obtained by replacing the covariance of each Gaussian component in the traditional Gaussian mixture regression. After iterating through and calculating all Gaussian components, the covariance matrix is ​​obtained, which consists of the covariance components corresponding to all Gaussian components in the attitude dimension. According to the covariance matrix The mean matrix of the attitude dimension is calculated to obtain the end-effector reference attitude trajectory;

[0023] S43: The reference pose trajectory distribution consists of the end-point reference position trajectory and the end-point reference attitude trajectory.

[0024] In step S5, the optimal hyperparameters of the kernel function are obtained by minimizing the root mean square error of the reproduced reference pose trajectory distribution.

[0025] The variational Bayesian method in S3 aims to automatically optimize the number of Gaussian components in the pose trajectory Gaussian mixture model. It uses the variational Bayesian method to learn a whole class of probabilistic models. For each probabilistic model, the posterior probability on the position trajectory data is calculated, and the parameters in the Gaussian mixture model are inferred using the variational Bayesian method. , as well as The maximum posterior distribution.

[0026] 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 then reproducing the Riemann space from the tangent space through the tangent space inverse mapping.

[0027] The method of parallel migration of attitude quaternion covariance in tangent space in S4 refers to the migration of attitude quaternion covariance in the tangent space of two attitude quaternions to ensure that their dispersion is correctly represented in the tangent space of different attitude quaternions.

[0028] The pose trajectory distribution kernelization representation method in S4 refers to kernelization representation of pose trajectory distribution by using kernel method and minimizing KL divergence between reference pose trajectory distribution and parameterized trajectory distribution;

[0029] For the kernelized representation of the attitude trajectory distribution, the reference attitude trajectory is first... Projected onto the starting point of the attitude trajectory To ensure the consistency of the attitude quaternion covariance across different tangent spaces, the covariance of the reference attitude trajectory will be used. From the mean The tangent space is parallelly migrated to the starting point of the attitude trajectory. The tangent space.

[0030] The root mean square error of the reproduced reference trajectory in S5 refers to the penalty for the difference between the reproduced trajectory and the reference trajectory by squaring the Euclidean distance in the position dimension and the rotational distance in the attitude dimension.

[0031] The reference pose trajectory distribution update in S6 refers to using the job path points Update reference pose trajectory distribution For each job path point in Searching for Replace the trajectory points. The covariance of each work path point is a diagonal matrix, and the diagonal elements are a very small constant representing the accuracy of the trajectory adaptation to the work path points. The accuracy of the position dimension and the attitude dimension can be set separately.

[0032] The beneficial effects of this invention are:

[0033] This invention combines variational Bayesian and attitude quaternion tangent space mapping to construct a Gaussian mixture model of robot pose trajectory, which improves the accuracy of pose trajectory probability modeling while considering the geometric constraints of robot pose trajectory.

[0034] This invention considers the generalized generation of both position trajectory and attitude trajectory. The generated trajectory can adapt to changing work path points, making the robot's work path more flexible.

[0035] In summary, this invention achieves generalized generation of industrial robot pose trajectories that support variable work path points, thereby improving the flexibility of industrial robot work paths. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the method of the present invention.

[0037] Figure 2 The algorithm flowchart for constructing the Gaussian mixture model of robot pose trajectory in step S3 is shown.

[0038] Figure 3 This is a schematic diagram of the reference position trajectory calculated in step S4.

[0039] Figure 4 The algorithm flowchart for updating the reference pose trajectory distribution in step S62 is shown.

[0040] Figure 5 A schematic diagram of the end-effector pose trajectory generated in step S63. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings and specific examples.

[0042] like Figure 1 As shown, this embodiment of the invention uses the material barrel handling process in the nuclear waste treatment of industrial robots as an example for illustration, specifically including the following steps:

[0043] S1: Collect multiple sets of robot end-effector pose trajectories through drag-and-drop teaching, offline programming, virtual teaching, and other methods;

[0044] Robots include, but are not limited to, industrial robots, such as the ABB IRB6700 industrial robot.

[0045] In S1, drag teaching refers to the demonstration that the user completes a specific task by physically moving the robot joints in gravity compensation mode. At the same time, the robot's own sensors record the position and orientation of the end effector and the joint angles, or the robot's surrounding environment information (such as the position of obstacles, the status of other collaborative robots or the user) through an external vision system.

[0046] Offline programming in S1 refers to creating virtual working scenarios for robots in offline programming software such as ABB's RobotStudio and simulating and verifying the planned trajectories.

[0047] Virtual teaching in S1 refers to the convenient and intuitive control of a virtual environment or a remote robot through virtual reality, augmented reality, or other methods to collect the robot's end-effector pose trajectory.

[0048] S2: Align the time steps of multiple sets of robot end-effector pose trajectories to obtain the robot end-effector pose trajectory alignment set;

[0049] In S2, to address the issue that the time length of the robot end-effector pose trajectory varies each time, making it inconvenient to perform probabilistic modeling of the pose trajectory, a multidimensional dynamic time warping method is used to ensure the synchronization of position and attitude dimensions and satisfy the geometric constraints of attitude quaternions by setting the distance between trajectory points, thereby aligning the time steps of multiple sets of robot end-effector pose trajectories.

[0050] S2 specifically refers to:

[0051] S21: Convert the robot's end effector posture into a unit quaternion, which is a Riemannian manifold on a four-dimensional sphere;

[0052] S22: Calculate the distance between custom pose trajectory points. The distance between two trajectory points in the position and orientation dimensions are the Euclidean distance and the rotational distance, respectively. The rotational distance between the two orientation quaternions is the L2 regularization of the tangent space projection. ;

[0053] S23: Use dynamic programming to calculate the distance matrix consisting of the distances between each pair of data points;

[0054] S24: Solve for the alignment path between two trajectory sequences based on the distance matrix, given... Under the premise of reverse calculation To obtain the optimal alignment path ,satisfy ,in, Map coordinates to the trajectory points. and These represent the time series lengths of the two trajectory sequences, respectively. This indicates the length of the optimal alignment path.

[0055] S25: Align all pose trajectories with the target pose trajectory to obtain multiple sets of robot end effector pose trajectories after time alignment.

[0056] S3: Based on the robot end-effector pose trajectory alignment set, a Gaussian mixture model of robot pose trajectory is constructed by combining variational Bayesian and pose quaternion tangent space mapping methods.

[0057] like Figure 2 As shown, S3 specifically refers to:

[0058] S31: Based on the robot end-effector pose trajectory alignment set, a Gaussian mixture model of the robot position trajectory is constructed using the variational Bayesian method;

[0059] S32: Based on the robot end-effector pose trajectory alignment set, a Gaussian mixture model of robot pose trajectory is constructed using the pose quaternion tangent space mapping method.

[0060] S321: Through responsibility coefficient Weighted attitude trajectory points The mean of the Gaussian components is updated by projecting onto the mean tangent space. ;

[0061] S322: Mean of Gaussian components The dispersion of attitude quaternions in the tangent space is estimated as the covariance of the attitude quaternions. .

[0062] The robot pose trajectory Gaussian mixture model is composed of the robot position trajectory Gaussian mixture model and the robot attitude trajectory Gaussian mixture model.

[0063] S4: Based on the Gaussian mixture model of robot pose trajectory, combined with the method of parallel transfer of attitude quaternion covariance in tangent space and Gaussian mixture regression calculation, the reference pose trajectory distribution is obtained;

[0064] S4 specifically refers to:

[0065] S41: Calculate the end-effector reference position trajectory for the robot's bucket handling task based on the traditional Gaussian mixture regression method, such as... Figure 3 As shown;

[0066] S42: Parallel transfer the covariance of each Gaussian component in the Gaussian mixture model of the robot pose trajectory over the tangent space of the pose quaternions, and use the covariance in the parallel transferred tangent space. The covariance component of the pose dimension of the i-th Gaussian component is obtained by replacing the covariance of each Gaussian component in the traditional Gaussian mixture regression. After iterating through and calculating all Gaussian components, the covariance matrix is ​​obtained, which consists of the covariance components corresponding to all Gaussian components in the attitude dimension. According to the covariance matrix The mean matrix of the attitude dimension is calculated to obtain the end-effector reference attitude trajectory;

[0067] S43: The reference pose trajectory distribution consists of the end-effector reference position trajectory and the end-effector reference attitude trajectory. Each trajectory point in the calculated reference pose trajectory distribution is... It includes time points. Mean of terminal position Variance of terminal location Mean of end attitude and end attitude variance .

[0068] S5: Perform optimal kernel representation on the reference pose trajectory distribution to obtain the optimal hyperparameters of the kernel function;

[0069] S5 specifically refers to:

[0070] S51: The end position trajectory distribution is kernelized by using the kernel method and minimizing the KL divergence between the reference pose trajectory distribution and the parameterized trajectory distribution;

[0071] S52: For the kernelized representation of the end-effector attitude trajectory distribution, firstly, the end-effector reference attitude trajectory... Projected onto the starting point of the attitude trajectory The tangent space;

[0072] S53: To ensure the consistency of attitude quaternion covariance across different tangent spaces, the covariance of the end-point reference attitude trajectory is... From the mean The tangent space is parallelly migrated to the starting point of the attitude trajectory. The tangent space;

[0073] S54: Select a suitable kernel function from Gaussian kernel, Cauchy kernel, and periodic kernel for the robot's bucket handling task;

[0074] S55: The root mean square error is used to penalize the difference between the reproduced trajectory and the reference trajectory by squaring the Euclidean distance in the position dimension and the rotational distance in the attitude dimension.

[0075] S56: The optimal hyperparameters of the kernel function are solved by minimizing the root mean square error of the reproduced end-point reference pose trajectory.

[0076] S6: Update the reference pose trajectory distribution based on the robot's work path points, and generate the robot end pose trajectory adapted to the work path points by kernelizing the current reference pose trajectory distribution according to the optimal hyperparameters of the kernel function.

[0077] S6 specifically refers to:

[0078] S61: Set operation path points according to the changes in the robot's bucket handling operation scenario. Each operation path point is represented as follows: ,in, This indicates the expected time to travel through the work path points. and These represent the desired end-effector position and orientation, respectively. and These represent the covariance of the adaptive operation path points. The covariance is a diagonal matrix, and the diagonal elements are very small constants, representing the accuracy of the trajectory adaptation operation path points. The accuracy of the position dimension and the attitude dimension can be set separately.

[0079] S62: For each desired job path point, according to Figure 4 The process updates the reference pose trajectory distribution.

[0080] S63: Generate the robot end-effector pose trajectory adapted to the work path points using the kernel representation of the pose trajectory distribution. The generated end-effector pose trajectory of the industrial robot is as follows: Figure 5 As shown.

[0081] In the process of handling nuclear waste using industrial robots, the robot's workspace is limited, and the initial pose of the bucket on the worktable can change. There are specific requirements for the pose of the bucket placed on the worktable support, and it is also necessary to ensure that the robot does not collide with its surroundings. Therefore, the robot needs to have flexible operation capabilities. The end effector trajectory generated by this invention enables the industrial robot to adapt to the changing work path points in the bucket handling process, improving the flexibility of its work path.

[0082] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for generalizing the generation of pose trajectories for industrial robots that supports variable work path points, characterized in that, Includes the following steps: S1: Collect multiple sets of robot end-effector pose trajectories; S2: Align the time steps of multiple sets of robot end-effector pose trajectories to obtain the robot end-effector pose trajectory alignment set; S3: Based on the robot end-effector pose trajectory alignment set, a Gaussian mixture model of robot pose trajectory is constructed by combining variational Bayesian and pose quaternion tangent space mapping methods. Specifically, S3 is: S31: Based on the robot end-effector pose trajectory alignment set, a Gaussian mixture model of the robot position trajectory is constructed using the variational Bayesian method; S32: Based on the robot end-effector pose trajectory alignment set, a Gaussian mixture model of robot pose trajectory is constructed using the pose quaternion tangent space mapping method. The robot pose trajectory Gaussian mixture model is composed of the robot position trajectory Gaussian mixture model and the robot attitude trajectory Gaussian mixture model. S4: Based on the Gaussian mixture model of robot pose trajectory, combined with the method of parallel transfer of attitude quaternion covariance in tangent space and Gaussian mixture regression calculation, the reference pose trajectory distribution is obtained; Specifically, S4 is: S41: Calculate the reference position trajectory of the robot end effector based on the Gaussian mixture regression method; S42: Parallel transfer the covariance of each Gaussian component in the Gaussian mixture model of the robot pose trajectory over the tangent space of the pose quaternions, and use the covariance in the parallel transferred tangent space. The covariance component of the pose dimension of the i-th Gaussian component is obtained by replacing the covariance of each Gaussian component in the traditional Gaussian mixture regression. After iterating through and calculating all Gaussian components, the covariance matrix is ​​obtained, which consists of the covariance components corresponding to the pose dimensions of all Gaussian components. According to the covariance matrix The mean matrix of the attitude dimension is calculated to obtain the end-effector reference attitude trajectory; S43: The reference pose trajectory distribution consists of the end-point reference position trajectory and the end-point reference attitude trajectory; S5: Perform optimal kernel representation on the reference pose trajectory distribution to obtain the optimal hyperparameters of the kernel function; S6: Update the reference pose trajectory distribution based on the robot's work path points, and generate the robot end pose trajectory adapted to the work path points by kernelizing the current reference pose trajectory distribution according to the optimal hyperparameters of the kernel function.

2. The method for generalizing the generation of pose trajectories of an industrial robot supporting variable work path points as described in claim 1, characterized in that, In S2, the multidimensional dynamic time warping method is used to ensure the synchronization of position and attitude dimensions and to satisfy the attitude quaternion geometric constraints by setting the distance between trajectory points, thereby aligning the time steps of multiple sets of robot end-effector pose trajectories.

3. The method for generalizing the generation of pose trajectories of an industrial robot supporting variable work path points as described in claim 1, characterized in that, In step S5, the optimal hyperparameters of the kernel function are obtained by minimizing the root mean square error of the reproduced reference pose trajectory distribution.

Images