Robot skill learning system and method
By improving the quaternion dynamic motion primitive model and combining it with a dynamic target subsystem driven by teleoperation and fifth-order polynomials, the problem of the inability to specify the endpoint angular velocity in robot skill learning was solved, and precise control of non-zero endpoint angular velocity was achieved, thereby improving the adaptability and accuracy of robot skill generalization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG SCI-TECH UNIV
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-15
AI Technical Summary
In existing technologies, quaternion dynamic motion primitives cannot independently specify the angular velocity at the moment the action ends, which makes it impossible for robots to meet the requirements of specific non-zero endpoint angular velocities in scenarios such as hitting, throwing, and kicking, thus limiting their application scope.
By improving the quaternion dynamic motion primitive model, combining the teleoperation isomorphic master and slave hands, and employing a dynamic target subsystem driven by unit quaternion encoding, numerical differentiation algorithm, and fifth-order polynomial, a skill generalization trajectory that satisfies the endpoint angular velocity constraint is generated, driving the robot to complete complex skill learning.
This technology enables robots to independently specify non-zero endpoint angular velocities in scenarios such as hitting a ball, improving the accuracy and generalization ability of skill learning, reducing hardware costs, and enhancing the adaptability and application scope of the model.
Smart Images
Figure CN122033945A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent control technology, and in particular to a robot skill learning system and method. Background Technology
[0002] Robot skill learning is the core of a robot's mastery of complex motion skills. Imitation learning achieves efficient skill transfer by replicating human demonstration trajectories. Dynamic motion primitives, as the mainstream framework in this field, can extract features from taught trajectories and perform generalization, and are widely used in robot grasping, assembly, and other scenarios. Since robot motion skills in actual operations involve not only positional changes but also complex end-effector posture changes, researchers have extended dynamic motion primitives to quaternion space to accurately encode and generalize posture trajectories, in order to adapt to the reproduction requirements of complex rotational motions.
[0003] The core of the existing technology is the quaternion dynamic motion primitive. This method uses unit quaternions to represent robot posture, effectively solving the problem of reproducing complex rotational motion. It breaks through the limitations of traditional dynamic motion primitives in posture modeling and has shown good application potential in tasks with high requirements for posture control accuracy, such as robot welding and surgical assistance.
[0004] However, traditional quaternion dynamic motion primitives have inherent defects. Their internal mechanism forces the endpoint angular velocity to converge to zero, making it impossible to independently specify the angular velocity at the moment the action ends. In scenarios such as hitting, throwing, and kicking, robots need to achieve a specific non-zero endpoint angular velocity. This defect makes it difficult for existing methods to meet such requirements, greatly limiting their application scope. Summary of the Invention
[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a robot skill learning method that can solve the technical problem that the existing technology has an inherent mechanism that forces the endpoint angular velocity to converge to zero, making it impossible to independently specify the angular velocity at the moment the action ends; while in scenarios such as hitting a ball, throwing, and kicking, the robot needs to achieve a specific non-zero endpoint angular velocity. This defect makes it difficult for the existing method to meet such requirements, which greatly limits its application scope.
[0006] In a first aspect, the present invention provides a robot skill learning system, comprising: a robot skill teaching module, a robot skill learning module, and a robot skill generalization module connected in sequence.
[0007] The robot skill teaching module specifically includes: a teleoperation isomorphic master hand, an isomorphic teleoperation unit, and a teleoperation slave hand connected in sequence.
[0008] The robot skill learning module specifically includes: a teaching skill trajectory data acquisition unit and an improved quaternion dynamic motion primitive skill learning unit connected in sequence.
[0009] The robot skill generalization module specifically includes: a target skill task setting unit, a robot skill trajectory planning unit, and a robot motion control unit connected in sequence.
[0010] The isomorphic teleoperation unit is used to establish motion mapping and communication connections between the isomorphic teleoperation master and the teleoperation slave.
[0011] The target skill task setting unit is used to set generalization task parameters, including: desired posture, desired angular velocity, desired angular acceleration, and desired motion duration.
[0012] The robot skill trajectory planning unit is used to generate skill generalization trajectories that satisfy the endpoint angular velocity constraint.
[0013] The robot motion control unit is used to convert skill generalization trajectories into control commands, and drive the teleoperated slave according to the control commands to complete dynamic tasks with strict requirements on the end state.
[0014] A second aspect of this invention provides a robot skill learning method, comprising: S1: Acquire robot trajectory data.
[0015] S2: Construct an improved quaternion dynamic motion primitive model through the robot skill learning module.
[0016] S3: Input the robot trajectory data into the improved quaternion dynamic motion primitive model and train the improved quaternion dynamic motion primitive model.
[0017] S4: Through the robot skill generalization module, set the generalization task parameters and build a dynamic target subsystem.
[0018] S5: Input the dynamic target subsystem into the trained improved quaternion dynamic motion primitive model to generate a skill generalization trajectory that satisfies the constraints.
[0019] S6: Drive the robot to move along a trajectory according to the skill generalization trajectory in order to complete the robot's skill learning.
[0020] A third aspect of the present invention provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the robot skill learning method as described in the second aspect.
[0021] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this embodiment of the invention, by improving the quaternion dynamic motion primitive model, the angular velocity and angular acceleration of the trajectory are calculated, and the hitting point corresponding to the peak angular velocity is determined, which is used to set the trajectory data of the robot. The angular velocity at the moment the action ends can be specified independently. At the same time, by improving the quaternion dynamic motion primitive model and the desired dynamic target subsystem, a skill generalization trajectory that satisfies the constraints is planned to drive the robot to perform trajectory movement, so that the robot can achieve a specific non-zero endpoint angular velocity. Attached Figure Description
[0022] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.
[0023] Figure 1 This is a schematic diagram of the structure of a robot skill learning system provided in an embodiment of the present invention.
[0024] Figure 2 This is a flowchart illustrating a robot skill learning method provided in an embodiment of the present invention.
[0025] Figure 3 This is a generalization result diagram provided by an embodiment of the present invention.
[0026] Explanation of reference numerals in the attached diagram: 1-Robot skill teaching module; 2-Robot skill learning module; 3-Robot skill generalization module. Detailed Implementation
[0027] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0028] The robot skill learning method provided by the present invention will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.
[0029] Reference manual attached Figure 1 The diagram shows a schematic representation of a robot skill learning system provided in an embodiment of the present invention.
[0030] This invention provides a robot skill learning system, comprising: a robot skill teaching module 1, a robot skill learning module 2, and a robot skill generalization module 3 connected in sequence.
[0031] The robot skill teaching module 1 specifically includes: a teleoperation isomorphic master hand, an isomorphic teleoperation unit, and a teleoperation slave hand connected in sequence.
[0032] The robot skill learning module 2 specifically includes: a teaching skill trajectory data acquisition unit and an improved quaternion dynamic motion primitive skill learning unit connected in sequence.
[0033] The robot skill generalization module 3 specifically includes: a target skill task setting unit, a robot skill trajectory planning unit, and a robot motion control unit connected in sequence.
[0034] The isomorphic teleoperation unit is used to establish motion mapping and communication connections between the isomorphic teleoperation master and the teleoperation slave.
[0035] The target skill task setting unit is used to set generalization task parameters, including: desired posture, desired angular velocity, desired angular acceleration, and desired motion duration.
[0036] The robot skill trajectory planning unit is used to generate skill generalization trajectories that satisfy the endpoint angular velocity constraint.
[0037] The robot motion control unit is used to convert skill generalization trajectories into control commands, and drive the teleoperated slave according to the control commands to complete dynamic tasks with strict requirements on the end state.
[0038] In one possible implementation, the teaching skill trajectory data acquisition unit is responsible for recording the end-effector posture trajectory of the teleoperated slave during the teaching process.
[0039] An improved quaternion dynamic motion primitive skill learning unit is used to construct an improved quaternion dynamic motion primitive model based on the end-effector posture trajectory.
[0040] The motion state of the teleoperation isomorphic master hand is transmitted to the teleoperation slave hand in real time.
[0041] The teleoperation slave hand follows the teleoperation isomorphic master hand movement and uses an end effector to complete the skill action demonstration.
[0042] In this embodiment of the invention, the real-time motion mapping of the teleoperated isomorphic master and slave hands ensures the accurate reproduction of the taught skill actions. High-fidelity end-effector posture trajectories are obtained by the taught skill trajectory data acquisition unit, and then the improved quaternion dynamic motion primitive skill learning unit is used to construct a targeted model. This not only ensures the consistency between the taught data and the actual motion, but also provides an algorithmic foundation for accurately representing the posture motion features for subsequent skill generalization, effectively improving the efficiency of skill learning and the generalization and adaptation capabilities of the model.
[0043] Reference manual attached Figure 2 The diagram shows a flowchart of a robot skill learning method provided by an embodiment of the present invention.
[0044] S1: Acquire robot trajectory data.
[0045] In one possible implementation, S1 specifically includes sub-steps S101 to S108: S101: Demonstrates racket swing skills in table tennis hitting tasks via remote control.
[0046] It should be noted that the demonstration of table tennis hitting skills by remotely controlling the hand can accurately reproduce the subtle movements such as the adjustment of the racket face angle and the change of the swing speed during the human hitting process. This provides high-fidelity teaching samples for subsequent end-effector trajectory acquisition, ensuring the accuracy and completeness of skill feature extraction.
[0047] S102: Obtain the attitude trajectory of the robot's end effector during the swing skill demonstration.
[0048] S103: The attitude trajectory is encoded using unit quaternion encoding.
[0049] Among them, unit quaternion encoding refers to a mathematical encoding method used to accurately represent rotational motion in three-dimensional space.
[0050] Specifically, the unit quaternion is encoded as follows: ; in, q Indicates the current posture. v Indicates the current attitude q The real part, u Indicates the current attitude q The imaginary part, S 3 R represents 4 A unit sphere in space.
[0051] ν ∈R is the scalar part. u ∈R 3 It is a vector part that satisfies the constraints. Each unit quaternion corresponds to a rotation transformation in the rotation group SO(3), but there is a double-covering relationship: quaternion and Indicates the same rotation.
[0052] It should be noted that using unit quaternions to encode the attitude trajectory can effectively avoid the gimbal lock problem when representing rotation with Euler angles, accurately describe the complex three-dimensional rotational motion of the end effector during the robot's swing, and provide continuous and unambiguous attitude data support for the subsequent construction and parameter training of the improved quaternion dynamic motion primitive model.
[0053] S104: Calculate the angular velocity trajectory corresponding to the encoded attitude trajectory using a numerical differentiation algorithm.
[0054] Among them, numerical differentiation algorithm refers to a numerical calculation method based on discrete sampled data to approximate the derivative (rate of change) of a function. Its core purpose is to solve the problem of rate of change in scenarios where the analytical expression of the variable cannot be obtained and only discrete sample points can be obtained.
[0055] Specifically, the numerical differentiation algorithm is: the central difference method based on Lie algebras.
[0056] Specifically, first consider the quaternions at adjacent time points. q i+1 and q i-1 Calculate the relative rotation by obtaining the rotation vector through the logarithmic mapping log(), and then divide by the time interval. The angular velocity trajectory is obtained by using forward and backward differencing for the boundary initiation and endpoint, respectively. The angular acceleration trajectory is calculated using the same method. ; Where log() represents the logarithmic function and arccos() represents the inverse cosine function. This represents the Euclidean norm.
[0057] It should be noted that by processing the attitude trajectory encoded by the unit quaternion using the numerical differentiation algorithm, the continuous angular velocity trajectory can be accurately extracted directly from the discrete attitude sampling data without the need for additional angular velocity sensors. This reduces hardware deployment costs and provides complete attitude and velocity feature data support for improving the training of the quaternion dynamic motion primitive model.
[0058] S105: Calculate the angular acceleration trajectory corresponding to the angular velocity trajectory using a numerical differentiation algorithm.
[0059] It should be noted that the numerical differentiation algorithm can accurately calculate the angular acceleration trajectory from the existing angular velocity trajectory data without the need to deploy additional dedicated sensors to collect angular acceleration information. This reduces hardware costs and data acquisition complexity, and can also supplement the full-dimensional motion feature data of attitude, angular velocity, and angular acceleration, providing complete and high-precision teaching samples to support the training of improved quaternion dynamic motion primitive models.
[0060] S106: Identify the hitting point corresponding to the peak angular velocity in the swing skill demonstration based on the angular velocity trajectory.
[0061] It should be noted that by accurately locating the key hitting points in the swing skill demonstration through the peak characteristics of the angular velocity trajectory, the subjective error of manually identifying the hitting point is avoided. At the same time, it provides an objective and accurate basis for the subsequent decomposition of the teaching trajectory into acceleration and deceleration primitives, ensuring the effectiveness and relevance of the training data for improving the quaternion dynamic motion primitive model.
[0062] S107: Based on the point of impact, the swing skill demonstration is divided into acceleration and deceleration primitives.
[0063] It should be noted that, based on the objective peak angular velocity of the hitting point, the swing trajectory is divided into two clearly defined primitives: acceleration and deceleration. This avoids the subjective bias of manual division and can specifically extract the differentiated motion features of the two stages. This provides structured data support for improving the phased training of the quaternion dynamic motion primitive model, while ensuring the smooth connection and precise control of the acceleration and deceleration trajectories during subsequent generalization.
[0064] S108: Determine the robot trajectory data based on the acceleration primitive, deceleration primitive, and angular acceleration trajectory.
[0065] It should be noted that by integrating the segmented trajectory features of acceleration and deceleration primitives with the full-dimensional motion data of angular acceleration trajectories, a structured and complete robot trajectory dataset is formed. This avoids the limitations of single-dimensional data and provides accurate and comprehensive input data for improving the phased training of the quaternion dynamic motion primitive model. At the same time, it ensures the controllability of motion parameters and the continuity of trajectory during subsequent skill generalization.
[0066] In this embodiment of the invention, through the full-process design of remote operation teaching, unit quaternion encoding, numerical differential operation and hitting point division primitives, the accurate acquisition and structured integration of multi-dimensional data of robot swing trajectory are realized, avoiding human identification bias and additional hardware deployment costs, and providing complete and high-fidelity attitude, angular velocity and angular acceleration staged teaching data support for improving the training of quaternion dynamic motion primitive model.
[0067] S2: Construct an improved quaternion dynamic motion primitive model through the robot skill learning module.
[0068] In one possible implementation, S2 specifically includes sub-steps S201 to S203: S201: The traditional quaternion dynamic motion primitive model is analyzed to obtain the traditional transformation subsystem.
[0069] Among them, the traditional conversion subsystem refers to one of the core components of the traditional quaternion dynamic motion primitive model. It is a functional module responsible for converting the output signal of the model regularization subsystem into the robot's posture trajectory and guiding the trajectory to converge toward the target posture.
[0070] Specifically, the formula for the traditional transformation subsystem is as follows: ; in, Indicates the total execution time of the trajectory. This represents the standardized angular acceleration. α z Represents the first gain constant. β z Let represent the second gain constant, and log represent the logarithmic mapping operator. q attr Represents a static target attractor. To represent quaternion multiplication, Indicates the current attitude q conjugate, η Represents the standardized angular velocity. f ( s ) represents the nonlinear forcing term. s Represents the phase variable. Indicates posture q Time derivative, express η pure quaternions, q Indicates the current posture. N Indicates the number of basis functions. Indicates the first i Gaussian radial basis functions c i Indicates the first i The center of each basis function h i Indicates the first i The width of the basis functions, where exp() represents the exponential function. This represents the current attitude error of the system, defined in the Lie algebra space.
[0071] It should be noted that by systematically analyzing the traditional quaternion dynamic motion primitive model, the core traditional transformation subsystem architecture and mapping logic are accurately extracted, providing a clear reference benchmark for the design of subsequent improved models. At the same time, it can clarify the root cause of the technical defect of the traditional model where the endpoint angular velocity is forced to converge to zero, laying a solid theoretical and data foundation for targeted optimization of the transformation subsystem.
[0072] S202: The static target attractor in the traditional transformation subsystem is replaced with a dynamic target subsystem driven by a fifth-order polynomial to obtain an improved transformation subsystem.
[0073] Specifically, the formula for improving the conversion subsystem is as follows: ; in, This indicates the current attitude of the dynamic target subsystem. Represents the angular velocity of the dynamic target subsystem. This represents the angular acceleration of the dynamic target subsystem.
[0074] Specifically, in the nonlinear forcing term f ( s In this context, an adaptive rotation transformation matrix is used. Alternative This ensures that the learned trajectory shape features remain unchanged under any new starting and ending poses. ; For the recurrence trajectory When generalizing from the taught trajectory to a new trajectory, The calculation formula is: ; in, Indicates the amplitude scaling factor. e demo The feature vector representing the demonstration trajectory. e new Let I represent the target feature vector of the new task, and let I represent the matrix. θ Indicates the rotation angle. k Represents the rotation axis vector, [ k ] x Represents the rotation axis vector k antisymmetric matrix, This represents the normalized vector of the demonstrative eigenvector. This represents the normalized vector of the target feature vector of the new task.
[0075] It should be noted that by replacing the static target attractor with a dynamic target subsystem driven by a fifth-order polynomial, the limitation of fixed targets in traditional models is broken. This allows for precise adaptation to task requirements with non-zero endpoint angular velocities in scenarios such as hitting a ball. At the same time, the smoothness of the fifth-order polynomial ensures the continuity of trajectory changes, giving the improved quaternion dynamic motion primitive model a flexible target adaptation capability and providing core technical support for subsequent skill generalization.
[0076] S203: Combining the improved transformation subsystem and the regularization subsystem, construct an improved quaternion dynamic motion primitive model.
[0077] The regularization subsystem refers to the core component of the traditional and improved quaternion dynamic motion primitive model, and is a key module that works in parallel and in synergy with the transformation subsystem.
[0078] Specifically, the formula for the regularization subsystem is as follows: ; in, State variables s The first derivative with respect to time, α s Represents the convergence gain coefficient. s Represents the phase variable, phase variable s The initial value of 1 decreases monotonically with time to 0, used to drive the nonlinear forcing term. f ( s ).
[0079] Furthermore, the transformation system for all degrees of freedom of the robot shares the same regularization system, driven by a unified phase variable 's', to ensure the temporal synchronization of the motion of each degree of freedom. A traditional position dynamic motion primitive model is used to model the position trajectory, while an improved quaternion dynamic motion primitive model is used to model the attitude trajectory.
[0080] It should be noted that by integrating the dynamic target adaptation capability of the improved transformation subsystem with the model stability constraints of the regularization subsystem, an improved quaternion dynamic motion primitive model is constructed. This model not only solves the core defect of the traditional model where the endpoint angular velocity is forced to converge to zero, but also ensures the smoothness of the trajectory and the controllability of the target parameters, providing high-performance algorithmic model support for the accurate generalization of complex skills such as robot ball hitting.
[0081] In this embodiment of the invention, by analyzing the defects of traditional models, replacing the static target attractor and integrating the regularization subsystem to construct an improved model, the technical limitation of the traditional model's endpoint angular velocity being forced to zero is overcome, giving the model the ability to adapt to non-zero endpoint angular velocity scenarios such as hitting a ball. At the same time, relying on the collaboration of multiple subsystems, the trajectory is ensured to be smooth and controllable, providing high-performance algorithm support for the generalization of robot skills.
[0082] S3: Input the robot trajectory data into the improved quaternion dynamic motion primitive model and train the improved quaternion dynamic motion primitive model.
[0083] In this embodiment of the invention, the improved quaternion dynamic motion primitive model is trained by relying on structured, multi-dimensional, phased robot trajectory data. This model can accurately capture the differentiated motion characteristics during the acceleration and deceleration phases of the swing skill, effectively avoiding the drawbacks of single training data in traditional models. It ensures that the trained model has the ability to adapt to scenarios with non-zero endpoint angular velocity, providing high-precision and highly adaptable algorithmic support for subsequent robot skill generalization.
[0084] Reference manual attached Figure 3 The diagram shows the generalization results provided by the embodiments of the present invention.
[0085] exist Figure 3 In the experiment, the position, linear velocity, attitude quaternion, and angular velocity of the teaching trajectory, generalized trajectory 1, generalized trajectory 2, and generalized trajectory 3 were measured respectively.
[0086] The angular velocity of the taught trajectory at the point of impact is [0.0031, -0.0155, 0.5311] rad / s. The final angular velocities of the three generalized attitude trajectories are [0.0002, -0.0000, 0.6047], [0.0002, -0.0000, 0.8059], and [0.0002, -0.0001, 1.0070] rad / s, which are consistent with the set value [0, 0, 0.6]. T [0,0,0.8] T and [0,0,1.0] T The error in rad / s can be kept within 10. -3 The value is on the order of rad / s, indicating that while maintaining the original taught motion trajectory characteristics, the final angular velocity at the point of impact (i.e., the endpoint of the acceleration primitive) precisely reaches the set value. This supports generalization to achieve different impact intensities and spin effects, based on the gain constant in the dynamic motion primitive parameters. α z and β z With adjustments, the error can be further reduced.
[0087] The robot skill generalization module allows you to set generalization task parameters and build a dynamic target system.
[0088] Specifically, the generalization task parameters include: the start time of the generalized trajectory, the total execution time of the generalized trajectory, the initial attitude, the desired endpoint attitude, the initial angular velocity, the desired endpoint angular velocity, the initial angular acceleration, and the desired endpoint angular acceleration.
[0089] In one possible implementation, S4 specifically includes sub-steps S401 to S406: S401: Constructing the relative rotation trajectory of a fifth-order polynomial structure based on Lie algebra space.
[0090] Among them, the Lie algebra space refers to the linear vector space corresponding to the Lie group (such as the special orthogonal group SO(3)) that describes three-dimensional rotations. It is the core mathematical tool connecting nonlinear rotation representations such as rotation matrices and quaternions with linear operations.
[0091] Specifically, the formula for the relative rotation trajectory is as follows: ; in, Represents the locus function of a fifth-order polynomial. Represents the fifth-order polynomial trajectory function with respect to time. t The first derivative, Represents the fifth-order polynomial trajectory function with respect to time. t The second derivative, ∑ denotes the summation function. a i Represents the fifth-order polynomial. b The coefficients of the summation variables, b This represents the variable to be summed.
[0092] It should be noted that the relative rotation trajectory, constructed based on the fifth-order polynomial structure in the Lie algebra space, realizes analytical interpolation of rotational motion in the tangent space. At the same time, relying on the characteristics of the fifth-order polynomial, the position, angular velocity and angular acceleration boundary conditions of the trajectory are precisely constrained, which meets the requirements of non-zero endpoint angular velocity in scenarios such as hitting the ball. This provides a high-precision benchmark for improving the generation of smooth and controllable rotation trajectories in the quaternion dynamic motion primitive model.
[0093] S402: Establish the real-time conversion relationship between the relative rotation trajectory and the dynamic target subsystem.
[0094] Specifically, the formula for real-time conversion of relationships is as follows: ; in, Indicates time t The expected posture quaternion, q init The quaternion representing the robot's initial pose. This represents the multiplication operation of quaternions. Indicates time t The expected angular velocity vector, Represents the left Jacobian matrix. Indicates time t The expected angular acceleration vector, It represents the first derivative of the left Jacobian matrix with respect to time.
[0095] in, The left Jacobian matrix is used to establish the correct mapping between Lie algebraic velocities and physical angular velocities in the base coordinate system. ; in, Represents the left Jacobian matrix for trajectory components The partial derivatives, Represents trajectory components The first derivative with respect to time, cos() represents the cosine function, and sin() represents the sine function. express A skew-symmetric matrix, when hour, .
[0096] For rotation vector r The exponential mapping operator is defined as: ; in, r This represents the rotation vector.
[0097] It should be noted that establishing a real-time conversion relationship between the relative rotation trajectory and the dynamic target subsystem enables the accurate transfer of smooth trajectory parameters in the Lie algebra space to the dynamic target module. This ensures that the dynamic target subsystem can call upon trajectory constraints that adapt to non-zero endpoint angular velocities in real time, thereby improving the real-time performance and accuracy of the model trajectory generation and providing core data interaction support for the stable reproduction of the robot's swinging skill.
[0098] S403: Set generalization task parameters through the robot skill generalization module.
[0099] Specifically, the formula for generalizing task parameters is as follows: ; in, t 0 indicates the start time of the trajectory. Indicates the time when the trajectory ends. q attr The quaternion representing the target attitude at the moment the trajectory terminates. ω final This represents the expected angular velocity vector at the moment the trajectory terminates. α init This represents the initial angular acceleration vector at the start of the trajectory. α final This represents the target angular acceleration vector at the moment the trajectory terminates.
[0100] It should be noted that by setting generalization task parameters on demand through the robot skill generalization module, key indicators such as hitting power, trajectory amplitude, and endpoint angular velocity can be flexibly configured in different scenarios. This breaks the limitations of fixed parameters in traditional models, allowing the improved quaternion dynamic motion primitive model to adapt to diverse swing skill generalization needs and enhance the model's scenario scalability and practical application value.
[0101] S404: Combining real-time transformation relations and generalized task parameters, the constraint equations of the dynamic subsystem are obtained.
[0102] Specifically, for the acceleration primitive, the initial attitude, angular velocity, and angular acceleration are set to be the same as the desired final attitude, angular velocity, and taught trajectory. Different desired final angular velocities are set for generalization. ω final [0,0,0.6] T [0,0,0.8] T and [0,0,1.0] T For deceleration primitives, the initial attitude, angular velocity, and angular acceleration are set to be the desired final attitude, angular velocity, and angular acceleration of acceleration primitives to facilitate smooth primitive connections. The desired final attitude, angular velocity, and angular acceleration are set to be the same as the taught trajectory.
[0103] Specifically, the formula for the constraint condition is as follows: ; in, Represents the initial attitude quaternion q init The conjugate quaternion, This represents the total execution time of the left Jacobian matrix on the trajectory. The inverse matrix.
[0104] It should be noted that by combining the real-time conversion relationship between the relative rotation trajectory and the dynamic target subsystem, and the generalized task parameter generation constraints, the constraints can have both the smoothness of the theoretical trajectory and the personalized requirements of the actual task. This accurately defines the parameter boundaries of the dynamic subsystem, ensuring that the trajectory generated by the improved model not only meets the non-zero endpoint angular velocity requirement, but also conforms to the specific indicators of different generalization scenarios, thereby improving the accuracy and reliability of skill generalization.
[0105] S405: Solve the constraint equations to obtain the coefficient vector of the fifth-order polynomial structure.
[0106] Specifically, the constraint equations are expanded and then solved using a linear solver: ; It should be noted that by solving the constraints of the fusion trajectory characteristics and generalization requirements, the coefficient vector of the fifth-order polynomial structure is accurately obtained. The abstract constraint requirements are transformed into quantization parameters that can be directly called by the dynamic target subsystem. The smoothness of the fifth-order polynomial is relied upon to ensure the continuity of the trajectory and derivative. At the same time, it is ensured that the generated relative rotation trajectory strictly matches the generalization task indicators such as the non-zero endpoint angular velocity, providing core quantization support for improving the model output of high-precision motion trajectories.
[0107] S406: Construct a dynamic target system based on real-time transformation relationships and coefficient vectors.
[0108] It should be noted that the dynamic target system is constructed based on the real-time conversion relationship between the relative rotation trajectory and the dynamic target subsystem and the quantization coefficient vector. This deeply integrates the theoretical trajectory characteristics with the generalization task constraints, ensuring that the trajectory output by the system has both smooth continuity and task adaptability. At the same time, it accurately meets the non-zero endpoint angular velocity requirements of scenarios such as hitting the ball, providing core functional support for improving the high-precision skill generalization of the model.
[0109] In this embodiment of the invention, a relative rotation trajectory is constructed based on the Lie algebra space, and constraints are generated by combining generalization task parameters and solving the coefficient vector. Finally, a dynamic target system adapted to task requirements is constructed, breaking through the limitations of fixed parameters in traditional models, accurately meeting the requirements of non-zero endpoint angular velocity in scenarios such as hitting the ball, and greatly improving the skill generalization accuracy and scenario adaptability of the improved model.
[0110] S5: Input the desired dynamic target subsystem into the trained improved quaternion dynamic motion primitive model to generate a skill generalization trajectory that satisfies the constraints.
[0111] In one possible implementation, S5 specifically involves: inputting the dynamic target subsystem into the transformation subsystem of the trained improved quaternion dynamic motion primitive model, and generating a skill generalization trajectory that satisfies the constraints through a numerical integration method.
[0112] Specifically, the formula for the numerical integration method is as follows: ; Where, Δ t Indicates the discrete time step. η ( t )express t The scaling angular velocity vector at time t, express t The first derivative of the scaling angular velocity vector at any given time.
[0113] In this embodiment of the invention, by combining the trained improved quaternion dynamic motion primitive model with the expected dynamic target subsystem to plan the generalized trajectory of the skill, the core constraint of non-zero endpoint angular velocity in scenarios such as hitting a ball can be accurately met, ensuring the smoothness and continuity of the trajectory and the controllability of motion parameters, and greatly improving the accuracy of robot complex skill generalization and its adaptability to practical applications.
[0114] S6: Drive the robot to perform trajectory movement according to the generalized trajectory of the skill, so as to complete the skill learning of the robot.
[0115] In this embodiment of the invention, the robot is driven to complete trajectory movement based on the skill generalization trajectory of precise planning. The algorithmic advantages of the improved model are transformed into the robot's actual skill capabilities. This not only verifies the effectiveness and generalization of the model, but also ensures that the robot can stably reproduce complex skills such as hitting a ball that require non-zero endpoint angular velocity, thus efficiently completing the skill learning task.
[0116] This invention provides a readable storage medium that stores a program or instructions on the storage medium. When the program or instructions are executed by a processor, they implement the steps of the robot skill learning method described above and achieve the same technical effect. To avoid repetition, this invention will not elaborate further.
[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.
Claims
1. A robot skill learning system, characterized in that, include: The robot skill teaching module (1), robot skill learning module (2), and robot skill generalization module (3) are connected in sequence. The robot skill teaching module (1) specifically includes: a teleoperation isomorphic master hand, an isomorphic teleoperation unit, and a teleoperation slave hand connected in sequence; The robot skill learning module (2) specifically includes: a teaching skill trajectory data acquisition unit and an improved quaternion dynamic motion primitive skill learning unit connected in sequence; The robot skill generalization module (3) specifically includes: a target skill task setting unit, a robot skill trajectory planning unit, and a robot motion control unit connected in sequence; The isomorphic teleoperation unit is used to establish motion mapping and communication connection between the isomorphic teleoperation master hand and the teleoperation slave hand; The target skill task setting unit is used to set generalization task parameters, wherein the parameters include: desired posture, desired angular velocity, desired angular acceleration, and desired motion duration; The robot skill trajectory planning unit is used to generate a skill generalization trajectory that satisfies the endpoint angular velocity constraint. The robot motion control unit is used to convert the skill generalization trajectory into control commands, and drive the teleoperation slave according to the control commands to complete a dynamic task with strict requirements on the endpoint state.
2. The robot skill learning system according to claim 1, characterized in that, The teaching skill trajectory data acquisition unit is responsible for recording the end-effector posture trajectory of the teleoperated slave during the teaching process. The improved quaternion dynamic motion primitive skill learning unit is used to construct an improved quaternion dynamic motion primitive model based on the end-effector posture trajectory. The motion state of the teleoperation isomorphic master hand is transmitted to the teleoperation slave hand in real time; The teleoperation slave hand follows the movement of the teleoperation isomorphic master hand and uses an end effector to complete the demonstration of the skill action.
3. A robot skill learning method, characterized in that, Applied to the robot skill learning system according to any one of claims 1 to 2; the method includes: S1: Acquire robot trajectory data; S2: Construct an improved quaternion dynamic motion primitive model through the robot skill learning module; S3: Input the robot trajectory data into the improved quaternion dynamic motion primitive model and train the improved quaternion dynamic motion primitive model; S4: Through the robot skill generalization module, set the generalization task parameters and construct a dynamic target system; S5: Input the desired dynamic target subsystem into the trained improved quaternion dynamic motion primitive model to generate a skill generalization trajectory that satisfies the constraints. S6: Drive the robot to perform trajectory movement according to the skill generalization trajectory to complete the robot's skill learning.
4. The robot skill learning method according to claim 3, characterized in that, S1 specifically includes: S101: Demonstrate the swinging skill in a table tennis hitting task by controlling the remote operation from the hand; S102: Obtain the attitude trajectory of the robot's end effector during the swing skill demonstration; S103: Encode the attitude trajectory using unit quaternion encoding; S104: Calculate the angular velocity trajectory corresponding to the encoded attitude trajectory using a numerical differentiation algorithm; S105: Calculate the angular acceleration trajectory corresponding to the angular velocity trajectory using a numerical differentiation algorithm; S106: Based on the angular velocity trajectory, identify the hitting point corresponding to the peak angular velocity in the swing skill demonstration; S107: Based on the hitting point, the swing skill demonstration is divided into acceleration and deceleration elements; S108: Determine the robot trajectory data based on the acceleration element, the deceleration element, and the angular acceleration trajectory.
5. The robot skill learning method according to claim 4, characterized in that, The numerical differentiation algorithm is specifically the central difference method based on Lie algebras.
6. The robot skill learning method according to claim 3, characterized in that, S2 specifically includes: S201: The traditional quaternion dynamic motion primitive model is analyzed to obtain the traditional transformation subsystem; S202: Replace the static target attractor in the traditional conversion subsystem with a dynamic target subsystem driven by a fifth-order polynomial to obtain an improved conversion subsystem; S203: Combining the improved transformation subsystem and the regularization subsystem, construct the improved quaternion dynamic motion primitive model.
7. The robot skill learning method according to claim 3, characterized in that, The generalization task parameters specifically include: the start time of the generalization trajectory, the total execution time of the generalization trajectory, the initial attitude, the desired endpoint attitude, the initial angular velocity, the desired endpoint angular velocity, the initial angular acceleration, and the desired endpoint angular acceleration.
8. The robot skill learning method according to claim 3, characterized in that, S4 specifically includes: S401: Constructing the relative rotation trajectory of a fifth-order polynomial structure based on Lie algebra space; S402: Establish the real-time conversion relationship between the relative rotation trajectory and the dynamic target subsystem; S403: Set the generalization task parameters through the robot skill generalization module; S404: Combining the real-time conversion relationship and the generalized task parameters, the constraint equations of the dynamic subsystem are obtained; S405: Solve the constraint equations to obtain the coefficient vector of the fifth-order polynomial structure; S406: Construct a dynamic target subsystem based on the real-time conversion relationship and the coefficient vector.
9. The robot skill learning method according to claim 3, characterized in that, Specifically, S5 involves inputting the dynamic target subsystem into the transformation subsystem of the trained improved quaternion dynamic motion primitive model, and generating a skill generalization trajectory that satisfies the constraints through numerical integration.
10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the robot skill learning method as described in any one of claims 3 to 9.