A method for generating trajectory correction of a robot arm by fusing dynamics constraints

Abstract

The application discloses a kind of fusion kinematics constraint's mechanical arm generation trajectory correction processing method, its method includes: the corresponding demonstration mechanical arm trajectory set of research mechanical arm is constructed, and initial trajectory of mechanical arm is generated using demonstration mechanical arm trajectory set by diffusion model;Corresponding trajectory correction processing model of research mechanical arm is constructed, and model training is carried out using demonstration mechanical arm trajectory set, and physical constraint module is with the same joint, same time torque true value data and torque prediction data as physical residual and constraint total physical residual minimum;Initial trajectory of mechanical arm is input trajectory correction processing model, and the trajectory after correction of mechanical arm is obtained by outputting.The application is by physical residual, joint loss function optimization and threshold correction mechanism to the generated trajectory dynamics consistency correction processing, effectively eliminates the dangerous trajectory such as joint torque limit physical constraint, improves the smoothness of trajectory and task execution success rate.

Description

Technical Field

[0001] This invention relates to the field of robot trajectory planning technology, and in particular to a method for correcting the trajectory of a robotic arm that incorporates dynamic constraints. Background Technology

[0002] In the field of robotic arm trajectory planning, data fitting and high-dimensional mapping capabilities based on machine learning have gradually become the development direction. Existing robotic arm trajectory planning mainly adopts data-driven trajectory generation methods. In recent years, diffusion policy has become a typical representative, modeling trajectory generation as a stepwise denoising process. By iteratively eliminating noise, it gradually recovers the high-dimensional motion sequence from the random distribution. Existing diffusion models (mainly using diffusion policy) have strong multimodal distribution modeling capabilities and high generation quality, and can express complex and diverse trajectory distributions. However, their data-driven generation process lacks explicit physical constraints, and the generated trajectory does not have dynamic rationality. This leads to the generation of trajectories that are inconsistent with the real dynamic characteristics of the robotic arm, specifically manifested as sudden changes in joint torque, severe acceleration distortion, and even violations of actuator physical limits, which seriously threaten the safety and task execution success rate during actual deployment. Hierarchical Diffusion Policy and Dynamics-Guided Diffusion Model attempt to introduce physical law terms into the loss function of the generative model. However, these methods cannot obtain key physical quantities such as joint torques, thus lacking sufficient verification of trajectory executability and dynamic rationality, and are difficult to flexibly integrate into existing trajectory generation frameworks of different types.

[0003] In summary, existing trajectory generation models (such as diffusion models) cannot strictly guarantee that the generated trajectory satisfies the constraints of the Lagrange dynamics equations of the robotic arm, and the physical feasibility of the generated trajectory lacks a hard guarantee. Most existing trajectory generation models (such as diffusion models) only output kinematic trajectories (such as joint positions and velocity sequences) without simultaneously generating the corresponding joint torque sequences. This makes it impossible for the model to learn the intrinsic coupling relationship between the trajectory and the driving torque, and it is impossible to evaluate the mechanical load and energy consumption of the trajectory. This can easily lead to physically inconsistent or even dynamically unstable trajectories, affecting the smoothness of the robotic arm's movement and the safety of its execution. In some cases, dangerous trajectories exceeding the joint torque limits may even be generated during the generation stage. Summary of the Invention

[0004] The purpose of this invention is to provide a method for correcting the trajectory generated by a robotic arm that integrates dynamic constraints. Using the residuals of the Lagrange partial differential equations of dynamics as explicit constraints, and through joint loss function optimization and threshold correction mechanisms, the trajectory generated by the diffusion model (or other existing trajectory generation models) is post-processed to correct the dynamic consistency. This effectively eliminates dangerous trajectories that exceed physical constraints such as joint torque limits, thereby improving the smoothness of the trajectory and the success rate of task execution.

[0005] The objective of this invention is achieved through the following technical solution: A method for trajectory correction processing of a robotic arm that integrates dynamic constraints, the method comprising: S1. Construct a set of teaching manipulator trajectories corresponding to the research manipulator, and use the diffusion model to generate the initial trajectory of the manipulator using the set of teaching manipulator trajectories. S2. Construct a trajectory correction processing model for the research robot arm. The trajectory correction processing model includes a robot arm motion simulation module, a physical constraint module, and a neural network correction module. The trajectory correction processing model is trained using a set of taught robot arm trajectories. The robot arm motion simulation module uses the taught robot arm trajectories in the set of taught robot arm trajectories to perform dynamic constraint motion simulation of the robot arm, obtains the angular position data and angular velocity data of each joint of the robot arm at time t, and calculates the true torque data. The neural network correction module is constructed with a Long Short-Term Memory (LSTM) network as the core architecture. The associated taught robot arm trajectory, angle data, and angular velocity data are input into the neural network correction module. The neural network correction module outputs the corrected trajectory of the robot arm and the torque prediction data. The physical constraint module uses the true torque data and the torque prediction data of the same joint at the same time as the physical residual and constrains the total physical residual to be minimized. S3. Input the initial trajectory of the robotic arm into the trajectory correction processing model, and output the corrected trajectory of the robotic arm.

[0006] To better realize the present invention, the teaching manipulator trajectory set includes the teaching manipulator original trajectory and the teaching manipulator corrected trajectory, which are interconnected.

[0007] Preferably, the true torque data of each joint at time t is obtained using the Newton-Euler recursive method. In time The true value of the torque is denoted as .

[0008] Preferably, the physical constraint module constructs the following Euler-Lagrange dynamic equations based on the Lagrange function: ,in For joints The true value of the torque at the current time. For joints The rate of change of momentum with time at the current time, L, is a Lagrangian function, and the Lagrangian function L is defined as the joint. The difference between total kinetic energy and total potential energy at the current time; For joints From the current angular position, For joints At the current time, angular velocity For joints The generalized momentum or conjugate momentum at the current time.

[0009] Preferably, the physical residual in the physical constraint module The expression is as follows: , Output joints for neural network correction module The predicted torque value at the current time. For joints The physical residual at the current time is used to obtain the total physical residual for all joints and all times. The physical constraint module constrains the total physical residual to be minimized.

[0010] Preferably, the trajectory correction processing model is constructed using the following joint loss function: , , Here, N represents the trajectory loss term, and N is the total number of samples. To correct the trajectory of the model's predicted trajectory coordinates on sample t or to teach the robotic arm the trajectory. The original trajectory of the teaching robotic arm for sample t, For physical constraint loss terms, These are the weight parameters.

[0011] Preferably, the trajectory correction processing model includes an intelligent threshold judgment mechanism module. This module defines the geometric difference between the corrected trajectory of the robotic arm and its initial trajectory at time t as the structural deviation. Set structural deviation threshold ,when Then the trajectory is corrected using a robotic arm; when If so, the initial trajectory of the robotic arm is preserved.

[0012] Preferably, the neural network correction module includes an input layer, multiple LSTM hidden layers, and an output layer. The input layer receives associated teaching robot arm trajectory, angle data, and angular velocity data, and the output layer adopts a fully connected structure to output the corrected robot arm trajectory and torque prediction data.

[0013] Preferably, the Newton-Euler recursive method calculates the driving torque of each joint sequentially by propagating from the end effector of the robotic arm toward the base through the inverse dynamics of the robotic arm.

[0014] Preferably, the diffusion model uses the teaching manipulator trajectory from the set of teaching manipulator trajectories to generate a new initial trajectory for the manipulator through forward noise addition and backward noise removal inference.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This invention performs dynamic consistency correction on the trajectory generated by the diffusion model (or other existing trajectory generation model) through physical residual, joint loss function optimization and threshold correction mechanism. This fundamentally suppresses the trajectory mutation, excessive curvature and acceleration jump problems common in unconstrained data-driven methods, effectively eliminates dangerous trajectories that exceed physical constraints such as joint torque limits, and improves the smoothness of the trajectory and the success rate of task execution. Experimental results show that in the Block-Pushing task, the trajectory curvature after processing by this invention is reduced by 64.85%, and the task success rate is increased from 0.86 to 0.92.

[0016] (2) By jointly modeling the end trajectory and joint torque, the present invention enables the model to output the corrected kinematic trajectory synchronously. The generated trajectory meets the physical constraints of the joint torque of the robotic arm, avoids problems such as sudden acceleration and torque overload, and realizes accurate evaluation of trajectory mechanical load and energy consumption. From the dynamic source, it ensures the executability of the trajectory and the safety of the robotic arm operation, solves the defect of existing methods that cannot fully verify the physical feasibility of the trajectory due to the lack of torque information, improves the executability and safety of the robotic arm control system, and ensures the physical feasibility of the generated trajectory.

[0017] (3) This invention uses an intelligent threshold judgment mechanism to adaptively balance dynamic correction and task semantic preservation with structural deviation. While ensuring the feasibility of trajectory dynamics, it effectively avoids the risk of destroying the original trajectory geometry and task meaning due to over-correction, and realizes safe and reliable processing from trajectory generation to dynamic correction. Attached Figure Description

[0018] Figure 1 This is a flowchart of the trajectory correction processing method for robotic arms according to the present invention; Figure 2 This is a schematic diagram illustrating the principle of generating the initial trajectory of the robotic arm using a diffusion model and combining it with the physical constraints of the present invention in an embodiment. Figure 3 This is a schematic diagram illustrating the principle of the trajectory correction processing model in the embodiment; Figure 4 This is a schematic diagram illustrating the principle of the intelligent threshold judgment mechanism module in the embodiment. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to embodiments: Example

[0020] like Figure 1 As shown, a method for correcting the trajectory of a robotic arm that integrates dynamic constraints is proposed. The method includes: S1. Construct a set of teaching manipulator trajectories corresponding to the research manipulator (the research manipulator in this invention is mainly applied to multi-joint serial manipulator systems, such as a six-axis manipulator). The set of teaching manipulator trajectory contains a large number of high-quality teaching manipulator trajectories, which are teaching trajectories that have been verified through practical application. Using a diffusion model (this invention mainly focuses on trajectory planning research based on diffusion models; in this embodiment, the diffusion model is an existing trajectory generation model, but it can also be replaced and applied to other trajectory generation models), the initial trajectory of the manipulator is generated using the set of teaching manipulator trajectories. The role of the trajectory generation model (including the diffusion model) is to generate new initial trajectories for the manipulator (equivalent to a trajectory generator). The technical problem this invention aims to solve is whether the initial trajectory of the manipulator satisfies motion constraints and whether it is smooth and executable. Figure 2 As shown, the diffusion model uses the teaching robot trajectory set to generate a new initial robot trajectory through a forward process of adding noise (gradually adding Gaussian noise to the original teaching robot trajectory to make it completely random noise data) and a reverse process of denoising inference (the diffusion model gradually denoises the noise data based on training and learning to generate a new trajectory).

[0021] S2. Construct a trajectory correction processing model corresponding to the research robot arm. The trajectory correction processing model includes a robot arm motion simulation module, a physical constraint module, and a neural network correction module. The trajectory correction processing model is trained using a set of taught robot arm trajectories. To better promote model training and improve model training accuracy, generalization ability, and robustness, the preferred set of taught robot arm trajectories in this invention includes interrelated original and corrected trajectories. The robot arm motion simulation module uses the taught robot arm trajectories from the set of taught robot arm trajectories to perform dynamic constraint motion simulation of the robot arm, obtaining the angular position data and angular velocity data of each joint of the robot arm at time t, and calculating the true torque data. The angular position data is represented as... , Let n be the number of joints (also representing the degrees of freedom of the robotic arm), and let angular velocity data be expressed as... To ensure data continuity and physical consistency, a fixed time step is used (i.e., the time interval t is fixed, and the sampling process obtains time-series data sequentially according to the time step), so that the state changes at adjacent times can truly reflect the dynamic behavior of the robotic arm; joint angular acceleration It can be obtained through numerical differentiation using the central difference method, and the specific expression is as follows: The robotic arm motion simulation module can use existing motion simulation models that can output true torque data for each joint. The robotic arm trajectory, angular position data, angular velocity data, and true torque data of this invention are correlated to form a trajectory correction sample dataset. Other modules of the trajectory correction processing model use this trajectory correction sample dataset to perform physical constraints and trajectory correction processing on the model.

[0022] In some embodiments, the true torque data of each joint in the robotic arm motion simulation module at time t is obtained using the Newton-Euler recursive method. In time The true value of the torque is denoted as Thus, the true torque data is constructed. The Newton-Euler recursive method calculates the driving torque of each joint sequentially from the end effector of the robotic arm towards the base through inverse dynamics of the robotic arm (i.e., obtaining the true torque data). The Newton-Euler recursive method calculates the velocity and acceleration of each joint link through forward recursion, and then calculates the force and torque of each joint through reverse recursion. Specifically, the forward recursion process starts from the base of the robotic arm and propagates towards the end effector, while the reverse recursion process starts from the end effector and propagates towards the base of the robotic arm, calculating the force and torque on each joint link sequentially, and finally obtaining the driving torque corresponding to each joint.

[0023] like Figure 3As shown, the neural network correction module is built with the Long Short-Term Memory (LSTM) network as its core architecture. The related teaching robot arm trajectory, angle data and angular velocity data are input into the neural network correction module (the robot arm motion simulation module of the trajectory correction processing model outputs the true torque data). The neural network correction module outputs the corrected trajectory and torque prediction data of the robot arm. The physical constraint module uses the true torque data and torque prediction data of the same joint and the same time as the physical residual and constrains the total physical residual to be minimized. Specifically, the neural network correction module includes an input layer, multiple LSTM hidden layers, and an output layer. The input layer receives associated teaching robot arm trajectory, angle data, and angular velocity data. The multiple LSTM hidden layers include multiple LSTM units, each containing three gating structures: a forget gate, an input gate, and an output gate. These control the forgetting of historical information, the writing of new information, and the output of the hidden state, respectively. The forget gate determines how much of the unit state from the previous time step is retained in the current time step, the input gate determines how much of the input information from the current time step is stored in the unit state, and the output gate determines how much of the current unit state is output as the hidden state. In this embodiment, through the gating mechanism of the LSTM unit, the LSTM network can adaptively learn the dependencies at different time scales, effectively capturing the inertial effects and coupling characteristics in the robot arm's motion. The output layer uses a fully connected structure to output the corrected robot arm trajectory and torque prediction data. The network-predicted joint torque is compared with the joint state obtained through kinematics. Substitute into the Euler-Lagrange dynamics equations and calculate the physical residuals Furthermore, the physical residuals participate as physical constraint terms in the joint loss function for optimizing and updating network parameters. In this embodiment, the physical residuals in the physical constraint module... The expression is as follows: , Output joints for neural network correction module The predicted torque value at the current time. For joints The physical residual at the current time is used to obtain the total physical residual for all joints and all times. The physical constraint module constrains the total physical residual to be minimized.

[0024] In some embodiments, the physical constraint module constructs the following Euler-Lagrange dynamic equations based on the Lagrange function: ,in For joints The true value of the torque at the current time (the current time is the recorded time, for example, time t). For joints The rate of change of momentum with time at the current time, L, is a Lagrangian function, and the Lagrangian function L is defined as the joint. The difference between the total kinetic energy (i.e., the driving kinetic energy of the robotic arm) and the total potential energy (mainly gravitational potential energy) at the current time. For joints From the current angular position, For joints At the current time, angular velocity For joints The generalized momentum or conjugate momentum at the current time. Ideally, when the network output fully satisfies the robot arm's dynamics constraints, its physical residual should approach a zero vector, i.e., the physical residual... This embodiment uses automatic differentiation technology to calculate the partial derivatives of the Lagrangian function with respect to generalized coordinates and generalized velocities, thereby applying physical constraints in the continuous domain.

[0025] S3. Input the initial trajectory of the robotic arm into the trajectory correction processing model, and output the corrected trajectory of the robotic arm. To simultaneously achieve data fitting accuracy and dynamic consistency during model training, the preferred trajectory correction processing model of this invention constructs the following joint loss function: , , Here, N represents the trajectory loss term, and N is the total number of samples. To correct the trajectory of the model's predicted trajectory coordinates on sample t or to teach the robotic arm the trajectory. The original trajectory of the teaching robotic arm for sample t (in this embodiment, the set of teaching robotic arm trajectories is used as samples for model training. During training, constraints are applied through a joint loss function to ensure the basic accuracy of the model in trajectory prediction, which is an indispensable component of the supervised learning framework). For physical constraint loss terms, . This is a weighting parameter; when the task has high requirements for dynamic consistency, it can be increased. The value of ; conversely, it can be appropriately reduced; The L2 norm squared represents the difference. By minimizing the joint loss function, the model is guided to converge to a solution space that conforms to dynamic constraints while learning the data distribution characteristics. This invention addresses the shortcomings of purely data-driven trajectory generation methods, such as non-physical abrupt changes and insufficient trajectory smoothness. By introducing the residual term of the Lagrange dynamics partial differential equation as an explicit physical constraint, the initial trajectory generated by the model is corrected, ensuring that the generated trajectory always conforms to the dynamic motion laws of the robotic arm. This fundamentally suppresses the trajectory abrupt changes, excessive curvature, and acceleration jumps common in unconstrained data-driven methods. Experimental results show that in the Block-Pushing (Transformer) task (a classic standard test task in robot imitation learning / visual motion control), the trajectory curvature processed by this invention is reduced by 64.85%, and the curvature acceleration index also shows a significant downward trend, proving that the generated trajectory is smoother in geometry and more stable in dynamic characteristics.

[0026] To prevent over-correction of the original trajectory from affecting the task execution success rate, the trajectory correction processing model of this invention also includes an intelligent threshold judgment mechanism module, such as... Figure 4 As shown, the geometric difference between the corrected trajectory of the robotic arm and its initial trajectory at time t, obtained by the intelligent threshold judgment mechanism module, is defined as the structural deviation. Set structural deviation threshold ,when (Indicating that the corrected trajectory does not differ significantly from the original trajectory), then the corrected trajectory is used with the robotic arm; when (If the corrected trajectory indicates that it has significantly deviated from the geometric structure and task semantics of the original trajectory, directly adopting it may affect the task success rate), then the initial trajectory of the robotic arm is retained. The intelligent threshold judgment mechanism module of this invention effectively avoids the risk of destroying task semantics due to over-correction while ensuring dynamic feasibility.

[0027] This invention relates to trajectory correction processing for industrial robot arms, specifically correcting the consistency of robot arm trajectory dynamics. The specific workflow is as follows: The trajectory correction processing model is trained using a set of taught robot arm trajectories. A diffusion model or other trajectory generation model outputs the robot arm trajectory to be executed based on current observation conditions (such as target pose and obstacle information obtained from visual sensors). The robot arm motion simulation module uses the taught robot arm trajectory from the set of taught robot arm trajectories to perform dynamic constraint motion simulation of the robot arm, obtaining the angular position data and angular velocity data of each joint of the robot arm at time t, and calculating the true torque data. The neural network correction module is built with a Long Short-Term Memory (LSTM) network as its core architecture. The associated taught robot arm trajectory, angle data, and angular velocity data are input into the neural network correction module, which outputs the corrected robot arm trajectory and torque prediction data. The physical constraint module uses the true torque data and torque prediction data of the same joint at the same time as the physical residual and constrains the total physical residual to be minimized. After receiving the initial trajectory of the robot arm, the trajectory correction processing model outputs the corrected trajectory of the robot arm. The trajectory correction processing model has an intelligent threshold judgment mechanism module. This module defines the geometric difference between the corrected trajectory of the robotic arm and its initial trajectory at time t as the structural deviation. Set structural deviation threshold ,when Then the trajectory is corrected using a robotic arm; when The initial trajectory of the robotic arm is preserved. This invention continuously verifies and corrects the generated trajectory as necessary, ensuring that the end effector of the robotic arm can complete tasks such as grasping, assembly, and welding according to a smooth and dynamically feasible trajectory. This invention solves the problems of trajectory physical infeasibility, abrupt motion changes, and joint oscillations caused by the lack of dynamic constraints in existing data-driven robot trajectory generation methods. It can be seamlessly integrated into existing industrial robot control architectures without any structural modifications or retraining of the original trajectory generation model, significantly reducing the technical threshold for introducing dynamic constraints. Experimental results show that in the Block-Pushing task, the trajectory curvature processed by this invention is reduced by 64.85%, and the task success rate is improved from 0.86 to 0.92. In multiple complex operation tasks on the MimicGen dataset, this invention also demonstrates stable trajectory smoothing capabilities. Validated on various tasks, the method proposed in this invention not only guarantees a 1% increase in task success rate on average, but also effectively improves trajectory smoothness, with an average reduction of 9.73% in curvature and curvature acceleration.

[0028] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for correcting the trajectory of a robotic arm that integrates dynamic constraints, characterized in that: The methods include: S1. Construct a set of teaching manipulator trajectories corresponding to the research manipulator, and use the diffusion model to generate the initial trajectory of the manipulator using the set of teaching manipulator trajectories. S2. Construct a trajectory correction processing model for the research robot arm. The trajectory correction processing model includes a robot arm motion simulation module, a physical constraint module, and a neural network correction module. The trajectory correction processing model is trained using a set of taught robot arm trajectories. The robot arm motion simulation module uses the taught robot arm trajectories in the set of taught robot arm trajectories to perform dynamic constraint motion simulation of the robot arm, obtains the angular position data and angular velocity data of each joint of the robot arm at time t, and calculates the true torque data. The neural network correction module is constructed with a Long Short-Term Memory (LSTM) network as the core architecture. The associated taught robot arm trajectory, angle data, and angular velocity data are input into the neural network correction module. The neural network correction module outputs the corrected trajectory of the robot arm and the torque prediction data. The physical constraint module uses the true torque data and the torque prediction data of the same joint at the same time as the physical residual and constrains the total physical residual to be minimized. S3. Input the initial trajectory of the robotic arm into the trajectory correction processing model, and output the corrected trajectory of the robotic arm.

2. The method for correcting the trajectory of a robotic arm based on fused dynamic constraints according to claim 1, characterized in that: The teaching manipulator trajectory set includes the original teaching manipulator trajectory and the corrected teaching manipulator trajectory, which are interconnected.

3. The method for correcting the trajectory of a robotic arm based on fused dynamic constraints according to claim 1, characterized in that: The true torque data of each joint at time t were obtained using the Newton-Euler recursive method. In time The true value of the torque is denoted as .

4. The method for correcting the trajectory of a robotic arm based on fused dynamic constraints according to claim 1, characterized in that: The physical constraint module constructs the following Euler-Lagrange dynamic equations based on the Lagrange function: ,in For joints The true value of the torque at the current time. For joints The rate of change of momentum with time at the current time, L, is a Lagrangian function, and the Lagrangian function L is defined as the joint. The difference between total kinetic energy and total potential energy at the current time; For joints From the current angular position, For joints At the current time, angular velocity For joints The generalized momentum or conjugate momentum at the current time.

5. The method for correcting the trajectory of a robotic arm based on fused dynamic constraints according to claim 4, characterized in that: The physical residual in the physical constraint module The expression is as follows: , Output joints for neural network correction module The predicted torque value at the current time. For joints The physical residual at the current time is used to obtain the total physical residual for all joints and all times. The physical constraint module constrains the total physical residual to be minimized.

6. The method for correcting the trajectory of a robotic arm based on fused dynamic constraints according to claim 2, characterized in that: The trajectory correction processing model is constructed using the following joint loss function: , , Here, N represents the trajectory loss term, and N is the total number of samples. To correct the trajectory of the model's predicted trajectory coordinates on sample t or to teach the robotic arm the trajectory. The original trajectory of the teaching robotic arm for sample t, For physical constraint loss terms, These are the weight parameters.

7. The method for correcting the trajectory of a robotic arm based on fused dynamic constraints according to claim 1, characterized in that: The trajectory correction processing model includes an intelligent threshold judgment mechanism module. This module defines the geometric difference between the corrected trajectory of the robotic arm and its initial trajectory at time t as the structural deviation. Set structural deviation threshold ,when Then the trajectory is corrected using a robotic arm; when If so, the initial trajectory of the robotic arm is preserved.

8. The method for correcting the trajectory of a robotic arm based on fused dynamic constraints according to claim 1, characterized in that: The neural network correction module includes an input layer, multiple LSTM hidden layers, and an output layer. The input layer receives associated teaching robot arm trajectory, angle data, and angular velocity data. The output layer uses a fully connected structure to output the corrected robot arm trajectory and torque prediction data.

9. The method for correcting the trajectory of a robotic arm based on fused dynamic constraints according to claim 3, characterized in that: The Newton-Euler recursive method calculates the driving torque of each joint sequentially by propagating from the end effector of the robotic arm toward the base through the inverse dynamics of the robotic arm.

10. The method for correcting the trajectory of a robotic arm based on fused dynamic constraints according to claim 1, characterized in that: The diffusion model uses the teaching manipulator trajectory set to generate a new initial trajectory for the manipulator through forward noise addition and backward noise removal inference.

Images