Task-driven active and passive fusion mechanical arm motion evaluation method and system

Abstract

The invention discloses a task-driven active and passive fusion mechanical arm motion evaluation method and system, and the method comprises the steps: guiding a mechanical arm task, and collecting a joint angle data sequence; performing time alignment and smoothing processing to obtain an ideal demonstration track; inputting into a dynamic motion primitive model for learning to obtain model parameters; generalizing a new joint angle sequence conforming to the trajectory shape; a real-time control instruction is generated by taking the real-time control instruction as a reference track and adopting a reinforcement learning controller, and the mechanical arm is driven to track the reference track; and a multi-dimensional evaluation index system is constructed, and quantitative evaluation is conducted on the motion performance, the energy consumption efficiency and the intelligent level of task execution of the mechanical arm. According to the method, the robustness and the self-adaptive capability of the mechanical arm in different task scenes are remarkably improved, so that the efficiency and the safety of automatic operation are improved while stable operation is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical arm motion control, and particularly relates to a task-driven active-passive fusion mechanical arm motion evaluation method and system. BACKGROUND

[0002] Traditional methods based on manual programming and rule setting are difficult to cope with variable real tasks, while task-driven learning methods such as imitation learning and reinforcement learning provide new possibilities for robot autonomous operation. Reinforcement learning can learn from trial and error, and imitation learning can learn from human demonstrations. By recording the behavior data of humans performing tasks and analyzing it, robots can learn how to plan and execute actions in complex environments, thereby improving autonomous decision-making capabilities and broadening application scenarios.

[0003] In the motion control of a mechanical arm, existing methods still face significant challenges: first, although some research attempts to combine active and passive control methods, most of them only stay at simple control layer switching or weighted fusion, and lack a unified evaluation and learning framework oriented to task completion quality; second, the adaptability of current methods to environmental changes and task phase transitions is insufficient, and their control strategy parameters mostly rely on expert experience presetting or offline learning, which is difficult to optimize and adjust online based on real-time task execution feedback, resulting in limited robustness and intelligence in dynamic and unstructured environments. SUMMARY

[0004] The present application overcomes the deficiencies of the prior art, and to achieve the above purposes, a task-driven active-passive fusion mechanical arm motion evaluation method and system are used to solve the problems raised in the background art.

[0005] A task-driven active-passive fusion mechanical arm motion evaluation method, comprising the following steps: Step S1, through manual demonstration by an operator, guiding the mechanical arm to perform a task, and using an optical motion capture system to collect multiple demonstration joint angle data sequences; time aligning and smoothing the collected multiple joint angle data sequences to obtain an ideal demonstration trajectory representing task characteristics; Step S2, inputting the ideal demonstration trajectory into a dynamic motion primitive model for learning, and encoding to obtain model parameters representing the trajectory shape; inputting a new motion starting point and target point into the learned dynamic motion primitive model to generalize a new joint angle sequence conforming to the trajectory shape; Step S3, using the generalized new joint angle sequence as a reference trajectory, using a reinforcement learning controller based on a deep deterministic policy gradient algorithm to generate real-time control instructions, driving the mechanical arm to track the reference trajectory, and dynamically compensating and optimizing errors in the tracking process; Step S4, a multi-dimensional evaluation index system is constructed to quantitatively evaluate the motion performance, energy consumption efficiency and intelligent level of the robot arm in performing the task, wherein the motion performance at least includes end trajectory tracking error and joint angle overshoot, and the intelligent level includes a generalization ability score of unknown tasks.

[0006] As a further scheme of the present application: the specific process of collecting joint angle data by the optical motion capture system in step S1 includes: Marking points are arranged on each link and gripper of the robot arm to construct a 6-DOF rigid body model of the robot arm; The optical motion capture system is globally calibrated to define a world coordinate system, a robot arm base coordinate system and an end tool coordinate system; The spatial pose of the marking points is tracked by the motion capture system to inversely solve the joint angle values of each joint of the robot arm.

[0007] As a further scheme of the present application: the time alignment and smoothing processing of the plurality of joint angle data sequences in step S1 includes the following specific steps: The dynamic time warping algorithm is used to align the time steps, and the Gaussian mixture model and Gaussian mixture regression algorithm are combined to filter out data noise and fluctuations to generate the ideal demonstration trajectory.

[0008] As a further scheme of the present application: the specific steps in step S2 include: The dynamic motion primitive algorithm is trained by a curve to record the curve shape and reproduce and generalize according to requirements; Wherein, the dynamic motion primitive DMP is composed of a second-order dynamic system with self-stability, i.e. a spring damping system, and a force term learning trajectory shape The formula is:

[0009] Wherein, is a time constant, is a system state, and are the first derivative and the second derivative of , respectively, and are normal numbers, is a target point of the encoded trajectory, is a time-independent quantity; so that the DMP can model multiple-DOF trajectories simultaneously, and comes from a first-order regular system and satisfies: is a constant, is a nonlinear forcing term for learning trajectory shape, expressed as:

[0010] wherein, is the initial point of the encoding curve, is the number of basis functions, is the basis function, is the weight value of the basis function, and the entire process uses a Gaussian basis function, and the formula is:

[0011] wherein, and is the center and width of the first Gaussian basis function; the DMP can record the shape of the ideal demonstration curve, and the input ideal demonstration curve is , the nonlinear term obtained is:

[0012] The weight parameter is calculated by using the optimization method LWR ; through the obtained weight parameter , the new starting point and the target point of the curve are input to the DMP, and the phase variable driving of the regular system obtains the specific position, velocity and acceleration values of the new curve consistent with the shape of the demonstration curve.

[0013] As a further scheme of the present application: the specific steps of the real-time trajectory control of the mechanical arm in step S3 include: Step S31, a control node is created for receiving DMP parameters and control instructions, and a subscription and publication topic is configured; Step S32, the reinforcement learning DDPG algorithm is implemented to optimize the motion strategy of the mechanical arm. According to the trajectory generated by the DMP model, the mechanical arm starts to perform actions for passive tracking, and automatically adjusts its state to keep on the demonstration trajectory; the DDPG strategy generates the optimal action in real time according to the current state, and at the same time, the closed-loop control system monitors the error between the actual position of the mechanical arm and the DMP target trajectory in real time, and performs dynamic compensation; Step S33, the state and trajectory deviation of the mechanical arm are monitored in real time through the visualization tool of ROS; according to the task demand and real-time feedback, the DMP model parameters or the exploration parameters of the DDPG strategy are adjusted to improve the control effect.

[0014] As a further scheme of the present application: the specific steps in step S4 further include: ​The comparison data of the actual motion trajectory and the reference trajectory of the robot arm, the joint state data and the time series data of the evaluation index are collected and displayed in real time through a visualization tool and a log interface of a robot operating system (ROS).

[0015] The technical solution of the second aspect is a robot arm control system for implementing the method described in any of the above, comprising: A data acquisition and processing module is configured to acquire demonstration data through an optical motion capture system and generate an ideal demonstration trajectory. An action learning and generalization module is configured to learn the ideal demonstration trajectory and generalize a new trajectory. A real-time control module is configured to implement real-time trajectory tracking and optimized control of the robot arm.

[0016] As a further solution of the present application, the data acquisition and processing module and the real-time control module communicate data through a VRPN protocol and a ROS system, and the real-time control module inversely solves the pose information into joint angles through a kinematics solver.

[0017] Compared with the prior art, the present application has the following technical effects: The technical solution described above addresses the above-mentioned pain points by proposing a task-driven active-passive fusion robot arm motion evaluation method. The method first acquires demonstration data using an optical motion capture device to form a high-quality demonstration data set. Then, the DMP and DDPG algorithms are innovatively fused for analysis. DMP is used to characterize and learn the demonstration trajectory to generate a generalizable and adjustable expected motion primitive. DDPG is introduced to use a reward function designed based on task completion as an optimization objective. Through online learning, intelligent adaptive adjustment of uncertainty is achieved. Finally, based on this fusion framework, a complete technical closed loop of "demonstration learning, active reinforcement, passive tracking, online evaluation, and continuous optimization" is formed, significantly improving the intelligence and success rate of task execution of the robot arm in a dynamic position environment.

[0018] Furthermore, by organically fusing the DMP demonstration model and the DDPG reinforcement learning algorithm, a closed-loop adaptive control mechanism is formed in the robot arm motion control. This mechanism not only addresses motion deviations and disturbances in real-time task scenarios, improving the accuracy and response speed of trajectory tracking, but also provides strong support for the continuous optimization of control strategies and system stability through comprehensive monitoring of system performance. BRIEF DESCRIPTION OF DRAWINGS

[0019] The specific embodiments of the present application are described in detail below with reference to the accompanying drawings: Figure 1A schematic diagram of steps of a mechanical arm motion evaluation method disclosed in embodiments of the present application; Figure 2 A flowchart of acquisition and processing of demonstration data in embodiments of the present application; Figure 3 A flowchart of DMP action learning, reproduction and generalization processes in embodiments of the present application; Figure 4 A flowchart of mechanical arm reproduction and generalization of demonstration actions in embodiments of the present application; Figure 5 A flowchart of steps of a DDPG algorithm in embodiments of the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0021] Please refer to Figure 1 In embodiments of the present application, a task-driven active-passive fusion mechanical arm motion evaluation method comprises the following steps: Step S1, a human operator manually demonstrates, guides the mechanical arm to perform a task, and collects joint angle data sequences of multiple demonstrations by using an optical motion capture system; time alignment and smoothing processing are performed on the collected multiple joint angle data sequences to obtain an ideal demonstration trajectory representing task characteristics; The specific process of collecting joint angle data by using the optical motion capture system in step S1 comprises: Marker points are arranged on each link and gripper of the mechanical arm, and a 6-DOF rigid body model of the mechanical arm is constructed; The optical motion capture system is globally calibrated, and a world coordinate system, a mechanical arm base coordinate system and an end tool coordinate system are defined; The spatial pose of the marker points is tracked by the motion capture system, and the joint angle values of each joint of the mechanical arm are obtained by inverse solution.

[0022] In this embodiment, time alignment and smoothing processing are performed on the multiple joint angle data sequences in step S1, and the specific steps are as follows: Dynamic time warping algorithm is used to align time steps, and Gaussian mixture model and Gaussian mixture regression algorithm are combined to filter data noise and fluctuations to generate the ideal demonstration trajectory.

[0023] In the specific embodiment, the steps for obtaining action primitives are as follows: The mechanical arm action is essentially a sequence of joint angle changes of each joint. An operator manually controls the mechanical arm to demonstrate a series of predetermined tasks or paths, and the mechanical arm data packet is parsed to obtain the angle change data of each joint in each demonstration. This demonstration method is simple to operate and has strong versatility in demonstrating the mechanical arm action in various tasks. The demonstrated action conforms to the mechanical arm motion law and can be directly used as the original demonstration data without establishing additional mapping.

[0024] During the entire demonstration data acquisition process, passive optical motion equipment is used to acquire the motion data of the mechanical arm and the gripper when grabbing objects. An OptiTrack system composed of 14 cameras and a computer is used. The 14 cameras are placed around the capture area, and the cameras are connected to the same OptiHub connector using dedicated USB data lines. The OptiHub communicates with the PC through Network. The lens emits infrared rays, which are reflected when they hit the marker points. The sensing matrix on the lens can receive the reflected infrared rays, and the intersection of multiple cameras can determine the coordinates of the marker points in three-dimensional space. An L-shaped calibration tool and a calibration frame are used for global calibration. The motion capture software calculates the accurate spatial position and angle of the lens, defines the world coordinate system W, and defines and calibrates the initial pose of the mechanical arm base coordinate system B and the end tool coordinate system T in the motion capture world coordinate system W. Multiple Marker markers are attached to the key positions of the mechanical arm links and the gripper to create a 6DoF rigid body model of the mechanical arm, and the motion of each joint is accurately tracked. The position and orientation of the marker points can be determined from the images captured by each camera, and 6 degrees of freedom information can be obtained. The collected marker data is first marked and trajectory filled, and then filtered to remove noise and high-frequency interference in the data.

[0025] The joint angle values of all joints of the mechanical arm and the gripper obtained in each demonstration and the corresponding time steps are used as the demonstration data of this action, and each demonstration is used as a sample sequence. The sample point is defined as where M is the number of demonstrations, j={1, 2,..., T} represents the time step, T is the total number of time steps in this demonstration, is the joint angle value of the s-th joint of the mechanical arm corresponding to the time step, and s is the joint number.

[0026] The amplitude and time of the operator dragging the mechanical arm to perform the action are different in multiple demonstrations, resulting in significant differences in the amplitude and time steps of the obtained joint angle data. At the same time, the adjustment process of the demonstrator during the demonstration process causes the obtained joint angle data to have large fluctuations and not be smooth. In order to obtain an ideal demonstration curve that reflects the action characteristics and is smooth from multiple demonstration data, a dynamic time warping algorithm combined with a Gaussian mixture model and a Gaussian mixture regression is used to process the demonstration data.

[0027] As Figure 2 shown in FIG. 1, a flow chart is shown to demonstrate the acquisition and processing of data; Step S2, input the ideal demonstration trajectory into the dynamic movement primitive model for learning, and encode to obtain model parameters representing the trajectory shape; by inputting a new movement starting point and target point into the learned dynamic movement primitive model, a new joint angle sequence conforming to the trajectory shape is generalized; In this embodiment, the action learning, reproduction and generalization steps are as follows: The ideal curve is encoded by using the dynamic movement primitive (DMP) algorithm, so that the robot arm can reproduce and generalize the action in the demonstration.

[0028] The dynamic movement primitive algorithm is trained by the curve, records the curve shape, and reproduces and generalizes according to the requirements, as shown in Figure 3 FIG. 2, a flow chart of the DMP action learning, reproduction and generalization process is shown.

[0029] The specific steps include: The dynamic movement primitive algorithm is trained by the curve, records the curve shape, and reproduces and generalizes according to the requirements; Among them, the dynamic movement primitive DMP is composed of a second-order dynamic system with self-stability, i.e. a spring damping system, and a force term for learning trajectory shape, and the formula is:

[0030] Among them, is the time constant, is the system state, and are the first derivative and the second derivative of , respectively, and are normal numbers, is the target point of the encoded trajectory, is a time-independent quantity; so that the DMP can model multiple degrees of freedom trajectories at the same time, from a first-order regular system and satisfy: is a constant, is a nonlinear forcing term for learning trajectory shape, expressed as:

[0031] Among them, is the initial point of the encoded curve, is the number of basis functions, is the basis function, The weight value of the base function is the weight value of the Gaussian base function, and the formula is:

[0032] wherein, and is the center and width of the m-th Gaussian base function; the DMP can record the shape of the ideal demonstration curve, and the input ideal demonstration curve is , and the nonlinear term obtained is:

[0033] The weight parameter is calculated by using the optimization method LWR ; the new starting point and the target point of the curve are input to the DMP by using the obtained weight parameter , and the phase variable driving of the regular system obtains the specific position, velocity and acceleration values of the new curve of the spring damping system consistent with the shape of the demonstration curve.

[0034] As shown in Figure 4 , the flowchart of the mechanical arm reproducing and generalizing the demonstration action is shown; A DMP is established for each joint of the mechanical arm, and the ideal demonstration curve of each joint obtained from the demonstration is coded into the weight value of the Gaussian base function, and the multiple joints modeled are kept synchronous by using a common regular system. When called, the initial and target joint angle values of the actual action are input to each joint DMP, and the joint angle value of each time step of each joint is generalized by the DMP, which is used to control the movement of each joint of the mechanical arm and the gripper to realize the reproduction and generalization of the demonstration action. The process is shown. The DMP algorithm converts the demonstration action into weight parameters for easy storage, and the algorithm directly encodes the joint angle data of the mechanical arm at the same time, which is suitable for demonstration actions of mechanical arms with different joint numbers, and has the ability of reproduction and generalization. Only the initial point and the target point are needed to directly control the joint angle curve of the mechanical arm to execute the task, and the ease of use is strong.

[0035] ​​Because the Optitrack motion capture system and the ROS system are completely independent, have different control methods and data formats, a platform needs to be established between the two systems to realize the conversion and real-time transmission of data between the two systems, and complete the acquisition of the six-degree-of-freedom pose information of the rigid body by the robot arm. The data of the centroid pose of the rigid body obtained by the optical motion capture system after DMP processing are sent to ROS in real time through the Vrpn (Virtual-reality peripheral network) protocol; a server and a client are built under ROS, the server converts the subscribed rigid body pose information to the coordinate system of the robot arm in real time by using a data processing method, and sends the converted information to the client in the form of a service; the joint angles are solved by using a kinematics solver, and the motion planning control is performed by using Moveit, so as to realize the real-time pose tracking of the rigid body by the robot arm.

[0036] Step S3, using the generalized new joint angle sequence as a reference trajectory, using a reinforcement learning controller based on a deep deterministic policy gradient algorithm to generate real-time control instructions, driving the robot arm to track the reference trajectory, and dynamically compensating and optimizing the error in the tracking process; Specifically, the real-time trajectory control step is: After DMP demonstration learning, a deep deterministic policy gradient (DDPG) algorithm is introduced for deep reinforcement learning, which is used to train the action network to output effective actions in a new state, and train the value network to make a more fair score for the actions in the new state, so as to enable the action network to receive effective feedback information and better iterate, and then output more efficient actions.

[0037] The DDPG algorithm is built according to the Actor-Critic algorithm, the action network takes corresponding actions according to the observed state of the environment and obtains corresponding rewards, the evaluation network scores the action network according to the actions taken and the state of the environment, the action network adjusts the strategy taken by itself according to the score obtained, and then the evaluation network updates itself according to the reward value obtained, so that the single-step training is completed.

[0038] When the policy gradient algorithm is used for network training, each action of the agent needs to be sampled according to a probability distribution, and then the best action is selected:

[0039] In the formula: is the probability distribution function. Therefore, the action finally obtained by the policy gradient is random, and the algorithm can achieve good results when the task is simple and the action space is small. However, when the task is complex and the action space is large, the use of the policy gradient cannot guarantee convergence to the optimal policy, and a large amount of data and computing resources are required for training. The DDPG algorithm uses a deterministic policy gradient algorithm, which simplifies the calculation by directly outputting a unique action value using a neural network. The specific calculation formula is:

[0040] In the formula: is the parameter of the network.

[0041] As shown in Figure 5 , the flowchart shows the steps of the DDPG algorithm; In this embodiment, the specific steps of real-time trajectory control of the robot arm in step S3 include: Step S31, create a control node for receiving DMP parameters and control instructions, and configure the subscription and publication topics at the same time; Step S32, implement the reinforcement learning DDPG algorithm to optimize the motion strategy of the robot arm. According to the trajectory generated by the DMP model, the robot arm starts to perform actions for passive tracking, and automatically adjusts its state to keep on the demonstration trajectory; the DDPG strategy generates the optimal action in real time according to the current state, while the closed-loop control system monitors the error between the actual position of the robot arm and the DMP target trajectory in real time, and performs dynamic compensation; Step S33, real-time monitoring of the state and trajectory deviation of the robot arm through the visualization tool of ROS; according to the task requirements and real-time feedback, adjust the exploration parameters of the DMP model parameters or the DDPG strategy to improve the control effect.

[0042] Step S4, construct a multi-dimensional evaluation index system to quantitatively evaluate the motion performance, energy efficiency and intelligence level of the robot arm in executing tasks; wherein the motion performance at least includes end trajectory tracking error and joint angle overshoot, and the intelligence level includes generalization ability score for unknown tasks.

[0043] In this embodiment, the specific steps in step S4 further include: Through the visualization tool and log interface of the robot operating system ROS, real-time acquisition and display of the comparison data of the actual motion trajectory and the reference trajectory of the robot arm, the joint state data and the time series data of the evaluation index.

[0044] In this embodiment, the robot arm motion evaluation step is: Based on the above definition and analysis model, the motion evaluation of the robot arm is driven by the task, which includes the following four steps: Task definition and scene description. Define the task scene, use the ideal trajectory obtained by DMP demonstration learning as the reference, determine the initial angle, speed and other basic states of each joint, and ensure the consistency of the starting conditions of the movement.

[0045] Evaluation index construction. Design a multi-dimensional comprehensive evaluation index that integrates motion performance, energy efficiency and intelligence level. 1) Motion performance: includes traditional indexes such as end trajectory tracking error (RMSE), joint angle overshoot, steady-state time and force control contact overshoot rate. 2) Energy efficiency: quantifies total energy consumption, average power and unit displacement energy cost. 3) Intelligence level: introduces task complexity index (TCI) and generalization score (GS), and uses dynamic Bayesian network to comprehensively perceive noise resistance, decision stability and execution efficiency to evaluate the adaptability of the system in new tasks and new environments.

[0046] Data acquisition and dynamic feedback. During the evaluation process, through the ROS visualization tool and the system built-in log interface, the following key data are displayed and recorded in real time: 1) the comparison curve of the actual motion trajectory of the robot arm and the DMP target trajectory; 2) state data (joint angle, speed and end position, etc.) and related time series performance indicators.

[0047] Evaluation test and verification. Design a cross-scene and cross-platform comparison and verification test to compare the method described in the invention with traditional control methods, form quantitative evaluation conclusions and performance optimization suggestions.

[0048] The technical solution of the second aspect: a robot arm control system for implementing the method described in any one of the above, comprising: A data acquisition and processing module for acquiring demonstration data through an optical motion capture system and generating an ideal demonstration trajectory; An action learning and generalization module for learning the ideal demonstration trajectory and generalizing new trajectories; A real-time control module for implementing real-time trajectory tracking and optimized control of the robot arm.

[0049] In this embodiment, the data acquisition and processing module and the real-time control module communicate data through the VRPN protocol and the ROS system; the real-time control module uses a kinematics solver to inverse the pose information to joint angles.

[0050] Beneficial effects: By combining the DMP teaching trajectory with the DDPG active reinforcement learning algorithm, high precision and fast response of the robot arm in real-time passive trajectory tracking are achieved. The system not only can automatically compensate the motion error in closed loop, but also can optimize the control strategy in real time through the task-driven evaluation mechanism, significantly improving the robustness and adaptive ability of the robot arm in different task scenarios, thereby ensuring stable operation while improving the efficiency and safety of automated operation.

[0051] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, the scope of the present application is defined by the appended claims and their equivalents, all of which should be included within the protection scope of the present application.

Claims

1. A task-driven method for evaluating the motion of a robotic arm that integrates active and passive motion, characterized in that, Includes the following steps: Step S1: The operator manually demonstrates and guides the robotic arm to perform the task, and uses an optical motion capture system to collect joint angle data sequences from multiple demonstrations; the collected joint angle data sequences are time-aligned and smoothed to obtain an ideal demonstration trajectory that characterizes the task. Step S2: Input the ideal demonstration trajectory into the dynamic motion primitive model for learning, and encode the model parameters that characterize the trajectory shape; by inputting new motion starting point and target point into the learned dynamic motion primitive model, a new joint angle sequence that conforms to the trajectory shape is generalized. Step S3: Using the generalized new joint angle sequence as a reference trajectory, a reinforcement learning controller based on a deep deterministic policy gradient algorithm is used to generate real-time control commands to drive the robotic arm to track the reference trajectory, and to dynamically compensate and optimize the error during the tracking process. Step S4: Construct a multi-dimensional evaluation index system to quantitatively evaluate the motion performance, energy efficiency, and intelligence level of the robotic arm in performing tasks; wherein, the motion performance includes at least the end-effector trajectory tracking error and joint angle overshoot, and the intelligence level includes a score for the ability to generalize to unknown tasks.

2. The task-driven active-passive fusion robotic arm motion evaluation method according to claim 1, characterized in that, The specific process of acquiring joint angle data using an optical motion capture system in step S1 includes: Marking points are set on each link and gripper of the robotic arm to construct a 6-DOF rigid body model of the robotic arm; Perform global calibration of the optical motion capture system, defining the world coordinate system, the robot arm base coordinate system, and the end effector coordinate system; The spatial pose of the marked points is tracked by the motion capture system, and the joint angle values ​​of each joint of the robotic arm are obtained by inverse kinematics.

3. The task-driven active-passive fusion robotic arm motion evaluation method according to claim 1, characterized in that, In step S1, time alignment and smoothing are performed on multiple joint angle data sequences. The specific steps are as follows: The time steps are aligned using a dynamic time warping algorithm, and Gaussian mixture model and Gaussian mixture regression algorithm are combined to filter out data noise and fluctuations, generating the ideal demonstration trajectory.

4. The task-driven active-passive fusion robotic arm motion evaluation method according to claim 1, characterized in that, The specific steps in step S2 include: The dynamic motion primitive algorithm is trained using curves, records the shape of the curves, and can be reproduced and generalized as needed; The Dynamic Motion Element (DMP) consists of a self-stabilizing second-order dynamic system, namely a spring-damped system, and a force term that learns the trajectory shape. The composition, and the formula are: in, It is a time constant. For system status, and They are respectively The first and second derivatives, and For positive integers, For the target point of the encoded trajectory, It is a quantity that is independent of time; This enables DMP to simultaneously model multi-degree-of-freedom trajectories from a first-order regularized system that satisfy: It is a constant. The nonlinear forcing term used to learn the trajectory shape is expressed as: in, This is the initial point of the coding curve. The number of basis functions. As basis functions, These are the weights of the basis functions. Throughout the process, the Gaussian basis function is used, and the formula is: in, and For the first The center and width of the Gaussian function; DMP can record the shape of the ideal demo curve, let the input ideal demo curve be... The obtained nonlinear term for: Calculate the weight parameters using the LWR optimization method. ; through the obtained weight parameters Input a new starting point for the curve into the DMP. and target point The regular system obtains phase variable driving force. The specific position, velocity, and acceleration values ​​that affect the new curve that matches the shape of the demonstration curve in the spring-damped system.

5. The task-driven active-passive fusion robotic arm motion evaluation method according to claim 1, characterized in that, The specific steps for real-time trajectory control of the robotic arm in step S3 include: Step S31: Create a control node to receive DMP parameters and control commands, and configure the topics for subscription and publication; Step S32: Implement the reinforcement learning DDPG algorithm to optimize the motion strategy of the robotic arm. Based on the trajectory generated by the DMP model, the robotic arm begins to perform actions for passive tracking, automatically adjusting its state to stay on the demonstration trajectory; the DDPG strategy generates the optimal action in real time based on the current state, while the closed-loop control system monitors the error between the actual position of the robotic arm and the DMP target trajectory in real time and performs dynamic compensation. Step S33: Monitor the status and trajectory deviation of the robotic arm in real time using ROS visualization tools; adjust the DMP model parameters or DDPG strategy exploration parameters according to task requirements and real-time feedback to improve control performance.

6. The task-driven active-passive fusion robotic arm motion evaluation method according to claim 1, characterized in that, The specific steps in step S4 also include: Using the visualization tools and log interface of the Robot Operating System (ROS), the system collects and displays in real time the comparison data between the actual motion trajectory of the robotic arm and the reference trajectory, joint status data, and time series data of the evaluation indicators.

7. A robotic arm control system for implementing the method according to any one of claims 1 to 7, characterized in that, include: The data acquisition and processing module is used to acquire demonstration data through the optical motion capture system and generate an ideal demonstration trajectory; The action learning and generalization module is used to learn the ideal demonstration trajectory and generalize to new trajectories; The real-time control module is used to achieve real-time trajectory tracking and optimized control of the robotic arm.

8. The control system of the task-driven active-passive fusion robotic arm motion evaluation method according to claim 7, characterized in that, The data acquisition and processing module communicates with the real-time control module via the VRPN protocol and the ROS system; the real-time control module uses a kinematic solver to inversely solve the pose information into joint angles.

Images