A neural network-based adaptive trajectory control system and method for a robot arm

Abstract

The application provides a neural network-based adaptive trajectory control system and method for a mechanical arm, which collects state information of a mechanical arm of a music robot in real time and loads performance task information; generates a performance expected trajectory for each mechanical arm node according to the performance task information, determines a trajectory control bias domain for each mechanical arm node according to the state information of the mechanical arm and the corresponding performance expected trajectory; performs bias approximation on the control parameters of the mechanical arm through a pre-trained neural network model to obtain trajectory dead zone bias compensation of the mechanical arm control model; generates a mechanical arm control torque vector according to the trajectory dead zone bias compensation of the mechanical arm and the trajectory control bias domain for each mechanical arm node, and performs adaptive trajectory control on the mechanical arm through the mechanical arm control torque vector. The technical solution provided by the application can optimize the error constraint mechanism and dead zone modeling compensation through a lightweight network structure to improve the trajectory control accuracy of the mechanical arm of the music robot.

Description

Technical Field

[0001] This application relates to the field of robotic arm control technology, and more specifically, to a neural network-based adaptive trajectory control system and method for robotic arms. Background Technology

[0002] Neural network-based adaptive trajectory control technology for robotic arms has demonstrated great potential in recent years for complex tasks involving high precision and multiple degrees of freedom, particularly in applications such as music robots where trajectory, timing, and dynamic control are extremely demanding. For example, in automated playing systems for instruments like the piano or guzheng, the robotic arm must not only achieve precise key-striking and string-plucking actions but also accurately control the timing, intensity, and dynamic performance of each note. These tasks place extremely high demands on the trajectory control of the robotic arm, which traditional rigid control methods struggle to meet. Against this backdrop, neural network-based adaptive trajectory control methods have become an important means of overcoming modeling uncertainties, actuator nonlinearities (such as dead zones and friction), and complex control environments.

[0003] However, musical performance is a high-frequency, rhythmically rigorous process, especially during high-speed continuous performances (such as playing fast passages or rolls), which places extremely high demands on the response speed of the control system. In existing technologies, neural networks require a large number of matrix operations when learning and updating weights online. This significantly increases the computational load, especially when the number of network nodes is large, leading to real-time control lag. Furthermore, because performance tasks involve frequent changes in trajectory and dynamics, the convergence speed of neural networks may be inconsistent under different operating conditions, and overfitting or instability may even occur, amplifying trajectory tracking errors and affecting the accuracy of note timing and performance expressiveness. Therefore, how to optimize error constraint mechanisms and dead-zone modeling compensation through lightweight network structures to improve the trajectory control accuracy of musical robotic arms is a challenge facing the industry. Summary of the Invention

[0004] This application provides a neural network-based adaptive trajectory control system and method for a robotic arm. It can optimize the error constraint mechanism and dead zone modeling compensation through a lightweight network structure to improve the trajectory control accuracy of the music robot robotic arm.

[0005] In a first aspect, this application provides a neural network-based adaptive trajectory control method for a robotic arm, comprising the following steps:

[0006] The system collects the status information of the robotic arm of the music robot in real time and loads the performance task information of the music robot.

[0007] Based on the performance task information, the expected performance trajectory of each robotic arm node of the music robot is generated, and the trajectory control bias domain of each robotic arm node of the music robot during the performance is determined according to the robotic arm state information and the corresponding expected performance trajectory.

[0008] A pre-trained neural network model is obtained, and the control parameters of the robotic arm of the music robot are approximated by bias through the neural network model, thereby obtaining the trajectory dead zone bias compensation of the robotic arm control model.

[0009] The mechanical arm control torque vector of the music robot is generated based on the dead zone offset compensation of the mechanical arm trajectory and the trajectory control offset domain of each mechanical arm node, and the mechanical arm is adaptively controlled by the mechanical arm control torque vector.

[0010] In some embodiments, the status information of the robotic arm of the music robot is collected in real time through an intelligent sensor array.

[0011] In some embodiments, the performance task information is a structured data set used to guide the music robot arm to complete performance actions.

[0012] In some embodiments, generating the expected performance trajectory of each robotic arm node of the music robot based on the performance task information specifically includes:

[0013] The start time and duration of each note in the performance task information are converted into corresponding action execution time windows;

[0014] Each note in the performance task information is mapped to a corresponding end-effector spatial position, and then the end-effector spatial smooth trajectory of each robotic arm node of the music robot is generated based on all the end-effector spatial positions.

[0015] The desired performance trajectory of each robotic arm node of the music robot is generated based on all action execution time windows and corresponding end-effector spatial smooth trajectories.

[0016] In some embodiments, determining the trajectory control bias domain of each robotic arm node of the music robot during the performance process based on the robotic arm state information and the corresponding expected performance trajectory specifically includes:

[0017] For each robotic arm node of the music robot during the performance, the real-time trajectory position and real-time node speed of the robotic arm node are extracted from the robotic arm state information.

[0018] The trajectory position offset coefficient of the robotic arm node is determined by the real-time trajectory position and the corresponding expected performance trajectory;

[0019] A virtual controller is designed based on the obstacle Lyapunov function, and the virtual trajectory speed of the robotic arm node is obtained according to the virtual controller and the corresponding expected performance trajectory.

[0020] The trajectory velocity offset coefficient of the robotic arm node is obtained from the real-time node velocity and the virtual trajectory velocity;

[0021] The trajectory control bias domain of the robotic arm node is constructed based on the trajectory position bias coefficient and the trajectory velocity bias coefficient, thereby obtaining the trajectory control bias domain of each robotic arm node of the music robot during the performance.

[0022] In some embodiments, the trajectory dead zone offset compensation of the robotic arm control model is obtained by biasing the control parameters of the music robot through the neural network model, thereby obtaining the trajectory dead zone offset compensation of the robotic arm control model.

[0023] The state information of the robotic arm and the control parameters of the robotic arm are used to form the network input vector;

[0024] The network input vector is input into the neural network model. The neural network model uses the Gaussian kernel function as a basis, automatically selects the center node of the receptive field, and continuously adjusts the network output in the control loop through the weight update law, thereby dynamically learning and approximating the nonlinear deviation term caused by the dead zone in the robotic arm control model.

[0025] The compensation term output by the neural network model is used as the trajectory dead zone offset compensation of the robotic arm control model.

[0026] In some embodiments, adaptive trajectory control of the robotic arm via the robotic arm control torque vector specifically includes:

[0027] The control torque vector of the robotic arm is applied in real time to the joint actuators corresponding to each node of the music robot. The joint actuators convert the corresponding trajectory control torque into the corresponding joint drive current and send it to the robotic arm joint through the interface, so that the robotic arm can complete the angle and position adjustment, thereby realizing the adaptive trajectory control of the music robot robotic arm.

[0028] Secondly, this application provides a neural network-based adaptive trajectory control system for a robotic arm, used to execute a neural network-based adaptive trajectory control method for a robotic arm, comprising:

[0029] The information acquisition module is used to collect the status information of the robotic arm of the music robot in real time and load the performance task information of the music robot.

[0030] The bias determination module is used to generate the expected performance trajectory of each robotic arm node of the music robot based on the performance task information, and to determine the trajectory control bias domain of each robotic arm node of the music robot during the performance process based on the robotic arm state information and the corresponding expected performance trajectory.

[0031] The bias compensation module is used to obtain a pre-trained neural network model, and to perform bias approximation on the control parameters of the robotic arm of the music robot through the neural network model, thereby obtaining the trajectory dead zone bias compensation of the robotic arm control model.

[0032] The trajectory control module is used to generate the control torque vector of the music robot's robotic arm based on the dead zone offset compensation of the robotic arm trajectory and the trajectory control offset domain of each robotic arm node, and to perform adaptive trajectory control of the robotic arm through the control torque vector.

[0033] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described neural network-based adaptive trajectory control method for robotic arms.

[0034] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned neural network-based adaptive trajectory control method for a robotic arm.

[0035] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0036] The adaptive trajectory control system and method for a robotic arm based on a neural network provided in this application involves: real-time acquisition of the robotic arm state information of a music robot and loading the music robot's performance task information; generation of the expected performance trajectory for each robotic arm node of the music robot based on the performance task information; determination of the trajectory control bias domain for each robotic arm node during performance based on the robotic arm state information and the corresponding expected performance trajectory; acquisition of a pre-trained neural network model; bias approximation of the robotic arm control parameters of the music robot using the neural network model to obtain the trajectory dead zone bias compensation of the robotic arm control model; generation of the robotic arm control torque vector based on the robotic arm trajectory dead zone bias compensation and the trajectory control bias domain of each robotic arm node; and adaptive trajectory control of the robotic arm using the robotic arm control torque vector.

[0037] Therefore, this application first generates the expected performance trajectory of each robotic arm node of the music robot based on the performance task information, and dynamically constructs the trajectory control bias domain by combining the actual robotic arm state information. This not only enables real-time quantification of trajectory execution error, but also provides clear and adjustable error boundary constraints for the subsequent controller, which helps to improve the system's response sensitivity to dynamic errors and enhance its adaptability to complex rhythms and fine movements. Then, by acquiring a pre-trained neural network model to approximate the parameters of the music robot's robotic arm with bias and achieving trajectory dead zone bias compensation, the adaptability of the control system to nonlinear uncertainties can be significantly improved. This is especially crucial in complex performance scenarios with input dead zones and dynamic modeling errors. By utilizing a lightweight network structure, the nonlinear deviation term in the system can be dynamically approximated while ensuring low computational overhead, achieving efficient real-time error compensation. Finally, by combining the robotic arm trajectory dead zone bias compensation with the trajectory control bias domain of each robotic arm node to generate a control torque vector, and implementing adaptive trajectory control accordingly, high-precision dynamic control of the music robot in complex performance scenarios can be achieved.

[0038] In summary, the technical solution adopted in this application can optimize the error constraint mechanism and dead zone modeling compensation through a lightweight network structure, thereby improving the trajectory control accuracy of the music robot arm. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is an exemplary flowchart of a neural network-based adaptive trajectory control method for a robotic arm, as shown in some embodiments of this application.

[0041] Figure 2 This is an exemplary flowchart illustrating the generation of the desired performance trajectory of each robotic arm node of a music robot according to some embodiments of this application;

[0042] Figure 3 This is a schematic diagram of the structure of a neural network-based adaptive trajectory control system for a robotic arm, as shown in some embodiments of this application.

[0043] Figure 4 This is a schematic diagram of the structure of a computer device that implements a neural network-based adaptive trajectory control method for a robotic arm, according to some embodiments of this application. Detailed Implementation

[0044] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0045] This application provides a neural network-based adaptive trajectory control system and method for a robotic arm. The core of this system involves real-time acquisition of the robotic arm's state information and loading the robot's performance task information. Based on the performance task information, the system generates the desired performance trajectory for each robotic arm node. The system then determines the trajectory control bias domain for each robotic arm node during performance based on the robotic arm's state information and the corresponding desired performance trajectory. A pre-trained neural network model is acquired, and the control parameters of the robotic arm are approximated using this model to obtain the trajectory dead zone bias compensation for the robotic arm control model. Based on the robotic arm trajectory dead zone bias compensation and the trajectory control bias domain for each robotic arm node, a robotic arm control torque vector is generated. This torque vector is then used to perform adaptive trajectory control on the robotic arm. This approach utilizes a lightweight network structure to optimize the error constraint mechanism and dead zone modeling compensation, thereby improving the trajectory control accuracy of the robotic arm.

[0046] To better understand the above technical solutions, a detailed description of the technical solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. (Refer to...) Figure 1 The figure is an exemplary flowchart of a neural network-based adaptive trajectory control method for a robotic arm according to some embodiments of this application. The figure mainly includes the following steps:

[0047] In step S101, the status information of the robotic arm of the music robot is collected in real time, and the performance task information of the music robot is loaded.

[0048] In some embodiments, the state information of the robotic arm of the music robot is collected in real time by an intelligent sensor array. Specifically, the intelligent sensor array integrated on each joint or end effector of the robotic arm can be used to sample and transmit the motion state of the robotic arm at the millisecond level, thereby obtaining the state information of the robotic arm of the music robot. The state information of the robotic arm includes the real-time trajectory position, real-time node speed and node torque of each robotic arm node, etc., and each joint on the robotic arm can be regarded as the robotic arm node.

[0049] It should be noted that, in this application, the performance task information is a structured data set used to guide the music robot arm to complete the performance action; in specific implementation, a specified performance segment can be read from a local trajectory database or an external interface through a task scheduler, thereby parsing the performance segment to obtain performance task information, such as the start time and duration of each note, the level of dynamics, and the performance type identifier.

[0050] In step S102, the expected performance trajectory of each robotic arm node of the music robot is generated based on the performance task information, and the trajectory control bias domain of each robotic arm node of the music robot during the performance is determined based on the robotic arm state information and the corresponding expected performance trajectory.

[0051] In some embodiments, reference Figure 2 As shown, this figure is an exemplary flowchart illustrating the generation of the desired performance trajectory of each robotic arm node of a music robot according to some embodiments of this application. In this embodiment, the generation of the desired performance trajectory of each robotic arm node of the music robot based on the performance task information can be achieved by the following steps:

[0052] In step S1021, the start time and duration of each note in the performance task information are converted into corresponding action execution time windows;

[0053] In step S1022, each note in the performance task information is mapped to a corresponding end-effector spatial position, and then the end-effector spatial smooth trajectory of each robotic arm node of the music robot is generated based on all the end-effector spatial positions.

[0054] In step S1023, the expected performance trajectory of each robotic arm node of the music robot is generated based on all action execution time windows and corresponding end-effector spatial smooth trajectories.

[0055] In practical implementation, firstly, the start time and duration of each note in the performance task information can be converted into corresponding action execution time windows. That is, the start time of the note is used as the lower bound of the action execution time window, and the sum of the start time and duration of the note is used as the upper bound. This method yields the action execution time window for each note. Secondly, each note in the performance task information can be mapped to its corresponding end-effector spatial position. This is achieved by establishing a note space mapping table based on the instrument layout (e.g., the position of keys or strings), thus mapping each note to a target point in three-dimensional space, i.e., the corresponding end-effector spatial position. The mapping rule is based on a static transformation matrix (e.g., simulation or calibration table) between the instrument's geometric dimensions, key / string numbers, and spatial coordinates. Thirdly, the smooth end-effector spatial trajectories of each robotic arm node of the music robot can be generated based on all end-effector spatial positions. This allows obtaining the spatial position sequence of each robotic arm node when the end of the arm is at each end-effector spatial position, and then using spline interpolation (cubic...)... The end-effector spatial smooth trajectory of each robotic arm node is constructed using a spline or minimum acceleration trajectory. Finally, the expected performance trajectory of each robotic arm node of the music robot can be generated based on all action execution time windows and the corresponding end-effector spatial smooth trajectory. That is, for each robotic arm node of the music robot, a timestamp can be added to the end-effector spatial smooth trajectory of the robotic arm node through all action execution time windows, and the end-effector spatial smooth trajectory can be converted into the angular trajectory of the robotic arm node, i.e. the expected performance trajectory, through the above method, the expected performance trajectory of each robotic arm node of the music robot can be generated.

[0056] In some embodiments, determining the trajectory control bias domain of each robotic arm node of the music robot during the performance process based on the robotic arm state information and the corresponding expected performance trajectory can be achieved in the following manner:

[0057] For each robotic arm node of the music robot during the performance, the real-time trajectory position and real-time node speed of the robotic arm node are extracted from the robotic arm state information.

[0058] The trajectory position offset coefficient of the robotic arm node is determined by the real-time trajectory position and the corresponding expected performance trajectory.

[0059] A virtual controller is designed based on the obstacle Lyapunov function, and the virtual trajectory speed of the robotic arm node is obtained according to the virtual controller and the corresponding expected performance trajectory.

[0060] The trajectory velocity offset coefficient of the robotic arm node is obtained from the real-time node velocity and the virtual trajectory velocity;

[0061] The trajectory control bias domain of the robotic arm node is constructed based on the trajectory position bias coefficient and the trajectory velocity bias coefficient, thereby obtaining the trajectory control bias domain of each robotic arm node of the music robot during the performance.

[0062] In practical implementation, firstly, for each robotic arm node during the performance, the real-time trajectory position and real-time node speed of the robotic arm node can be extracted from the robotic arm state information. Then, the trajectory position bias coefficient of the robotic arm node can be determined by the real-time trajectory position and the corresponding expected performance trajectory. The trajectory position bias coefficient represents the overall deviation between the real-time motion trajectory of the robotic arm node and the expected performance trajectory during the performance. The Euclidean distance between the real-time trajectory position corresponding to each note and its corresponding position in the expected performance trajectory can be calculated, and the standard deviation of all Euclidean distances can be used as the trajectory position bias coefficient of the robotic arm node. Secondly, a virtual controller can be designed based on an obstacle Lyapunov function. The virtual trajectory speed of the robotic arm node can then be obtained based on the virtual controller and the corresponding expected performance trajectory. When establishing a stable and adjustable control framework, a virtual controller based on an obstacle Lyapunov function is introduced. The role of this virtual controller is to dynamically construct an ideal virtual speed trajectory based on the current error state and expected performance indicators, guiding the actual speed to gradually approach the expected performance trajectory and ensuring that the trajectory error is correct. The error will not exceed the set boundary. The virtual trajectory speed not only considers the current expected trajectory of the performance, but also introduces the error convergence speed and adjustable performance parameters, so that the system can maintain stability and responsiveness even in rapidly changing or continuous performance scenarios. Furthermore, the trajectory speed bias coefficient of the robotic arm node can be obtained from the real-time node speed and the virtual trajectory speed. The trajectory speed bias coefficient represents the degree of overall difference between the real-time node speed and the virtual trajectory speed of the robotic arm node during the performance. The Euclidean distance between the real-time node speed corresponding to each note and the corresponding point in the virtual trajectory speed can be calculated, and the standard deviation of all Euclidean distances can be used as the trajectory speed bias coefficient of the robotic arm node. Finally, the trajectory control bias domain of the robotic arm node can be constructed based on the trajectory position bias coefficient and the trajectory speed bias coefficient. The trajectory control bias domain represents the overall deviation of the running trajectory under the current robotic arm control model. The dataset composed of the trajectory position bias coefficient and the trajectory speed bias coefficient can be used as the trajectory control bias domain of the robotic arm node. In the above way, the trajectory control bias domain of each robotic arm node of the music robot during the performance can be obtained.

[0063] It should be noted that generating the expected performance trajectory of each robotic arm node of the music robot based on the performance task information, and dynamically constructing the trajectory control bias domain in combination with the actual robotic arm state information, not only enables real-time quantification of trajectory execution error, but also provides clear and adjustable error boundary constraints for the subsequent controller. This helps to improve the system's response sensitivity to dynamic errors and enhance its adaptability to complex rhythms and fine movements (such as piano fast keys and guzheng glissando).

[0064] In step S103, a pre-trained neural network model is obtained, and the control parameters of the robotic arm of the music robot are approximated by bias through the neural network model, thereby obtaining the trajectory dead zone bias compensation of the robotic arm control model.

[0065] In some embodiments, obtaining a pre-trained neural network model can be achieved in the following ways:

[0066] The node structure of the neural network model is optimized by width learning, and the weights of the neural network model are updated using an adaptive update law, thereby obtaining a pre-trained neural network model.

[0067] In practical implementation, firstly, traditional neural networks (such as standard radial basis function neural networks) often suffer from low training efficiency and weak generalization ability when facing complex dynamic systems due to redundant nodes or rigid structures. This application optimizes the node structure of the neural network model through width learning. A width learning mechanism is introduced to dynamically optimize the node structure of the neural network. That is, during the learning process, based on the distribution of the input data, it is determined whether new nodes need to be added. New nodes are only created when the distance between a new input sample and the center of an existing node exceeds a certain threshold, and the current number of nodes does not exceed a preset upper limit. This avoids structural bloat. Furthermore, during each forward computation, the network only activates a few nodes closest to the current input (e.g., the first m nodes). The network involves multiple nodes in the computation, significantly reducing computational load and improving real-time performance. For activated nodes, the receptive field center and width parameters are adjusted according to new input samples, ensuring that the network structure remains highly coupled with the input features throughout the training process, thereby enhancing its ability to express the state of the robotic arm. Then, the weights of the neural network model can be updated using an adaptive update law. That is, based on the optimization of the network structure, an adaptive update law designed based on Lyapunov stability theory is used to achieve online updating and stable convergence of the neural network weights. During operation, the controller continuously collects the error between the actual state (position, velocity) and the desired state of the robotic arm nodes. This error information, together with the activation features of the network, serves as input to drive the weight adjustment equation, adjusting the connection weights of the neural network in real time. The update law design ensures error-oriented weight learning, and the Lyapunov function construction ensures that the update process does not lead to system instability, guaranteeing global convergence. Weight updates can continue until the network output stably approximates the trajectory dead zone bias term and the uncertain part of the model, thus forming an accurate and highly generalizable neural network model, i.e., obtaining the pre-trained neural network model.

[0068] In some embodiments, the trajectory dead zone offset compensation of the robotic arm control model is obtained by biasing the control parameters of the music robot through the neural network model, thereby obtaining the offset compensation of the robotic arm control model. Specifically, the following methods can be used:

[0069] The state information of the robotic arm and the control parameters of the robotic arm are used to form the network input vector;

[0070] The network input vector is input into the neural network model. The neural network model uses the Gaussian kernel function as a basis, automatically selects the center node of the receptive field, and continuously adjusts the network output in the control loop through the weight update law, thereby dynamically learning and approximating the nonlinear deviation term caused by the dead zone in the robotic arm control model.

[0071] The compensation term output by the neural network model is used as the trajectory dead zone offset compensation of the robotic arm control model.

[0072] In practical implementation, firstly, the current state information of the robotic arm of the music robot, together with the known control parameters of the robotic arm, constitutes the input vector of the neural network. The control parameters of the robotic arm include auxiliary information such as inertia matrix estimation, joint friction coefficient, and historical control torque. The robotic arm state information and control parameters are fused to form a high-dimensional feature vector, which serves as the input to the neural network to comprehensively reflect the current control state and dynamic changes. Then, the network input vector can be input into the neural network model. The neural network model uses a Gaussian kernel function as a basis and automatically selects the receptive field center nodes. That is, the neural network dynamically selects the m closest center nodes based on the distance between the current input sample and the existing node centers for activation and computation, improving operational efficiency. To avoid overfitting, the network output is continuously adjusted in the control loop using a weight update law. This involves continuously adjusting the network connection weights based on the control error-driven weight update law, allowing the neural network output to gradually approximate the trajectory deviation term caused by nonlinearities such as dead zone. This dynamically learns and approximates the nonlinear deviation term caused by dead zone in the robotic arm control model. Finally, the compensation term output by the neural network model can be used as the trajectory dead zone offset compensation for the robotic arm control model. This trajectory dead zone offset compensation represents the deviation caused by the dead zone effect in the current robotic arm trajectory control. If this trajectory dead zone offset compensation is applied to the actual control torque command generation module, dynamic correction of the dead zone error can be achieved, making the actual output torque more consistent with the ideal trajectory requirements.

[0073] It should be noted that by using a pre-trained neural network model to approximate the parameters of the music robot's robotic arm with bias and achieving trajectory dead zone bias compensation, the adaptability of the control system to nonlinear uncertainties can be significantly improved. This is especially crucial in complex performance scenarios with input dead zones and dynamic modeling errors. Utilizing a lightweight network structure, nonlinear deviation terms in the system are dynamically approximated while maintaining low computational overhead, achieving efficient real-time error compensation. Compared to traditional rigid control strategies, this adaptive compensation mechanism not only enhances the flexibility of trajectory error constraints but also effectively handles performance tasks with extremely high requirements for response speed and trajectory accuracy, such as rapid piano playing and guzheng glissando. It ensures that keystrokes are not delayed or misaligned, and that the playing force and rhythm are precisely consistent, thereby improving the performance fidelity and expressiveness of the music robot.

[0074] In step S104, the control torque vector of the music robot is generated based on the dead zone offset compensation of the robot trajectory and the trajectory control offset domain of each robot node, and the robot arm is adaptively controlled by the control torque vector.

[0075] In some embodiments, the generation of the robotic arm control torque vector for the music robot based on the dead zone bias of the robotic arm trajectory and the trajectory control bias domain of each robotic arm node can be specifically achieved in the following manner:

[0076] Design of a robotic arm controller based on Lyapunov stability theory;

[0077] The dead zone offset of the robotic arm trajectory and the trajectory control offset domain of each robotic arm node are input into the robotic arm controller to obtain the trajectory control torque of each robotic arm node.

[0078] The control torque vector of the music robot's robotic arm is generated by the trajectory control torque of all robotic arm nodes.

[0079] In practical implementation, firstly, a robotic arm controller can be designed based on Lyapunov stability theory. This involves constructing a robotic arm controller framework based on Lyapunov stability theory. This controller not only considers the feedback adjustment mechanism for conventional trajectory tracking errors but also integrates the modeling uncertainties and dead-zone nonlinear effects present in the model. The design goal of the robotic arm controller is to ensure that all closed-loop error systems are asymptotically stable under control input, meaning that all position and velocity errors can gradually converge within the bounded performance function, while avoiding overshoot or jitter, thus ensuring the stability and smoothness of the performance trajectory. Then, the dead-zone bias of the robotic arm trajectory and the trajectory control bias domain of each robotic arm node can be input into the robotic arm controller. Specifically, during the control execution phase, the dead-zone bias predicted by the neural network model and the trajectory control bias domain calculated in real-time by each robotic arm node are used as key parameters input to the designed robotic arm controller. In the device, the robotic arm controller comprehensively considers these inputs, namely, using trajectory dead zone bias information for nonlinear compensation, using trajectory control bias domain information to perform boundary constraints and corrections on the current joint error, and dynamically adjusting the control gain to achieve personalized trajectory repair for each robotic arm node, thereby outputting the control torque of each robotic arm node, i.e., obtaining the trajectory control torque of each robotic arm node. During this process, the robotic arm controller internally uses obstacle Lyapunov functions to construct a constraint domain to ensure that all biases are adjusted within a safe range, preventing actions from exceeding limits or keystroke deviations. This is particularly suitable for performance actions such as guzheng glissando and piano fast keys, which have extremely high requirements for trajectory continuity and trigger accuracy. Finally, the robotic arm control torque vector of the music robot can be generated from the trajectory control torques of all robotic arm nodes, that is, the feature vector formed by the trajectory control torques of all robotic arm nodes in the node order is used as the robotic arm control torque vector of the music robot.

[0080] In some embodiments, adaptive trajectory control of the robotic arm via the control torque vector can be achieved in the following manner:

[0081] The control torque vector of the robotic arm is applied in real time to the joint actuators corresponding to each node of the music robot. The joint actuators convert the corresponding trajectory control torque into the corresponding joint drive current and send it to the robotic arm joint through the interface, so that the robotic arm can complete the angle and position adjustment, thereby realizing the adaptive trajectory control of the music robot robotic arm.

[0082] In practical implementation, firstly, the control torque vector of the robotic arm can be applied in real time to the joint actuators corresponding to each node of the robotic arm of the music robot. That is, the trajectory control torque corresponding to each dimension of the control torque vector is applied in real time to the joint actuators corresponding to each node of the robotic arm of the music robot. Each trajectory control torque is bound to a specific joint motion axis to drive it to perform precise rotational or linear displacement movements. Then, the joint actuator internally converts the received trajectory control torque signal into a corresponding drive current command. Through the torque-current mapping model of the joint actuator (such as the torque constant inside a servo motor or direct drive motor), the efficient conversion from control quantity to physical driving force is achieved. Secondly, the converted drive current is transmitted through a high-speed real-time communication interface (such as USB, Ether)... The data is transmitted via CAT or CAN bus to the joint control module of the music robot arm. The interface supports low-latency, high-bandwidth data interaction, ensuring that each trajectory control command is issued and executed within milliseconds, meeting the response speed requirements of complex performance actions. Finally, upon receiving the control current, the joint control module drives the motor or servo mechanism to complete the corresponding angle and position adjustments, thereby accurately tracking the desired trajectory under the current control cycle. The adjustment process is continuous and driven by closed-loop feedback, which can correct trajectory errors caused by dead zone deviation, dynamic interference, or structural friction in real time, ensuring the accuracy and continuity of each keystroke, glissando, finger lift, and other actions. Through the above steps, the robot arm completes a response control to the target trajectory, thereby realizing the adaptive trajectory control of the music robot arm.

[0083] It should be noted that by combining dead zone bias compensation of the robotic arm trajectory with the trajectory control bias domain of each robotic arm node to generate a control torque vector, and implementing adaptive trajectory control accordingly, high-precision dynamic control of the music robot in complex performance scenarios can be achieved. This significantly enhances the model's ability to identify and compensate for nonlinear dead zone effects, model uncertainties, and real-time errors, making it particularly suitable for performance tasks that require extremely high responsiveness and trajectory accuracy, such as fast legato piano playing and glissando on the guzheng. By introducing a lightweight neural network structure, the system achieves online learning and rapid correction of dead zone modeling and error constraints under controllable computational overhead. This effectively reduces trajectory drift and action delay caused by model lag or input saturation, ensuring that each keystroke of the robotic arm is accurate, rhythmic, and consistent in force, improving the naturalness and musical expressiveness of the performance. This is a key control strategy for achieving high-fidelity anthropomorphic performance.

[0084] Therefore, this application first generates the expected performance trajectory of each robotic arm node of the music robot based on the performance task information, and dynamically constructs the trajectory control bias domain by combining the actual robotic arm state information. This not only enables real-time quantification of trajectory execution error, but also provides clear and adjustable error boundary constraints for the subsequent controller, which helps to improve the system's response sensitivity to dynamic errors and enhance its adaptability to complex rhythms and fine movements. Then, by acquiring a pre-trained neural network model to approximate the parameters of the music robot's robotic arm with bias and achieving trajectory dead zone bias compensation, the adaptability of the control system to nonlinear uncertainties can be significantly improved. This is especially crucial in complex performance scenarios with input dead zones and dynamic modeling errors. By utilizing a lightweight network structure, the nonlinear deviation term in the system can be dynamically approximated while ensuring low computational overhead, achieving efficient real-time error compensation. Finally, by combining the robotic arm trajectory dead zone bias compensation with the trajectory control bias domain of each robotic arm node to generate a control torque vector, and implementing adaptive trajectory control accordingly, high-precision dynamic control of the music robot in complex performance scenarios can be achieved.

[0085] In summary, the technical solution adopted in this application can optimize the error constraint mechanism and dead zone modeling compensation through a lightweight network structure, thereby improving the trajectory control accuracy of the music robot arm.

[0086] Furthermore, in another aspect of this application, in some embodiments, this application provides a neural network-based adaptive trajectory control system for a robotic arm, with reference to... Figure 3 The figure is a schematic diagram of the structure of a neural network-based adaptive trajectory control system for a robotic arm according to some embodiments of this application. The neural network-based adaptive trajectory control system for a robotic arm includes: an information acquisition module 201, an offset determination module 202, an offset compensation module 203, and a trajectory control module 204, which are described below:

[0087] The information acquisition module 201 is used to collect the status information of the robotic arm of the music robot in real time and load the performance task information of the music robot.

[0088] The bias determination module 202 is used to generate the expected performance trajectory of each robotic arm node of the music robot based on the performance task information, and to determine the trajectory control bias domain of each robotic arm node of the music robot during the performance process based on the robotic arm state information and the corresponding expected performance trajectory.

[0089] The bias compensation module 203 is used to obtain a pre-trained neural network model, and to perform bias approximation on the control parameters of the robotic arm of the music robot through the neural network model, thereby obtaining the trajectory dead zone bias compensation of the robotic arm control model.

[0090] The trajectory control module 204 is used to generate the control torque vector of the music robot's robotic arm based on the dead zone offset compensation of the robotic arm trajectory and the trajectory control offset domain of each robotic arm node, and to perform adaptive trajectory control of the robotic arm through the control torque vector of the robotic arm.

[0091] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described neural network-based adaptive trajectory control method for robotic arms.

[0092] In some embodiments, reference Figure 4 The figure is a schematic diagram of the structure of a computer device implementing a neural network-based adaptive trajectory control method for a robotic arm, according to some embodiments of this application. The neural network-based adaptive trajectory control method for a robotic arm in the above embodiments can... Figure 4 The computer device shown is used to implement this, and the computer device includes at least one processor 301, a communication bus 302, a memory 303, and at least one communication interface 304.

[0093] The processor 301 can be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more devices used to control the execution of the neural network-based adaptive trajectory control method for the robotic arm in this application.

[0094] The communication bus 302 can be used to transmit information between the aforementioned components.

[0095] The memory 303 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 303 may exist independently and be connected to the processor 301 via the communication bus 302. The memory 303 may also be integrated with the processor 301.

[0096] The memory 303 stores program code for executing the scheme of this application, and its execution is controlled by the processor 301. The processor 301 executes the program code stored in the memory 303. The program code may include one or more software modules. In the above embodiments, the determination of the neural network-based adaptive trajectory control method for the robotic arm can be implemented by the processor 301 and one or more software modules in the program code in the memory 303.

[0097] Communication interface 304 uses any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0098] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0099] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0100] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described neural network-based adaptive trajectory control method for robotic arms.

[0101] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0102] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A neural network-based adaptive trajectory control method for a robot arm, characterized by, The method comprises the following steps: Real-time acquisition of the state information of the mechanical arm of the music robot, and loading of the performance task information of the music robot; Generation of the performance expected trajectory of each mechanical arm node of the music robot according to the performance task information, determination of the trajectory control bias domain of each mechanical arm node of the music robot during the performance according to the state information of the mechanical arm and the corresponding performance expected trajectory; Obtaining of a pre-trained neural network model, bias approximation of the mechanical arm control parameters of the music robot through the neural network model, and further obtaining of the trajectory dead zone bias compensation of the mechanical arm control model; Generation of the mechanical arm control torque vector of the music robot according to the trajectory dead zone bias compensation of the mechanical arm and the trajectory control bias domain of each mechanical arm node, and adaptive trajectory control of the mechanical arm through the mechanical arm control torque vector. The generation of the performance expected trajectory of each mechanical arm node of the music robot according to the performance task information specifically comprises: Conversion of the start time and duration of each note in the performance task information into the corresponding action execution time window; Mapping of each note in the performance task information into the corresponding end space position, and further generation of the end space smooth trajectory of each mechanical arm node of the music robot according to all the end space positions; Generation of the performance expected trajectory of each mechanical arm node of the music robot according to all the action execution time windows and the corresponding end space smooth trajectory. The determination of the trajectory control bias domain of each mechanical arm node of the music robot during the performance according to the state information of the mechanical arm and the corresponding performance expected trajectory specifically comprises: For each mechanical arm node of the music robot during the performance, extraction of the real-time trajectory position and real-time node speed of the mechanical arm node from the state information of the mechanical arm; Determination of the trajectory position bias coefficient of the mechanical arm node through the real-time trajectory position and the corresponding performance expected trajectory; Design of a virtual controller based on a barrier Lyapunov function, and obtaining of the virtual trajectory speed of the mechanical arm node according to the virtual controller and the corresponding performance expected trajectory; Obtaining of the trajectory speed bias coefficient of the mechanical arm node from the real-time node speed and the virtual trajectory speed; Construction of the trajectory control bias domain of the mechanical arm node according to the trajectory position bias coefficient and the trajectory speed bias coefficient, and further obtaining of the trajectory control bias domain of each mechanical arm node of the music robot during the performance. The bias approximation of the mechanical arm control parameters of the music robot through the neural network model, and further obtaining of the trajectory dead zone bias compensation of the mechanical arm control model specifically comprise: Construction of a network input vector from the state information of the mechanical arm of the music robot and the mechanical arm control parameters; Input of the network input vector into the neural network model, automatic selection of a receptive field center node by the neural network model based on a Gaussian kernel function, continuous adjustment of the network output in the control loop through a weight update law, and further dynamic learning and approximation of the nonlinear deviation term caused by the dead zone in the mechanical arm control model; Taking the compensation term output by the neural network model as the trajectory dead zone bias compensation of the mechanical arm control model. 2.The neural network-based adaptive trajectory control method for a robot arm according to claim 1, wherein, The state information of the mechanical arm of the music robot is collected in real time through an intelligent sensor array. 3.The neural network-based adaptive trajectory control method for a robot arm according to claim 1, wherein, The performance task information is a structured data set for guiding the music robot mechanical arm to complete a performance action.

4. The neural network-based adaptive trajectory control method for a robot arm according to claim 1, wherein, The adaptive trajectory control of the mechanical arm through the mechanical arm control torque vector specifically includes: The mechanical arm control torque vector is applied to the joint actuator corresponding to each mechanical arm node of the music robot in real time, the joint actuator converts the corresponding trajectory control torque into corresponding joint driving current, and sends the joint driving current to the joint of the mechanical arm through an interface, so that the mechanical arm completes angle and position adjustment, and adaptive trajectory control of the music robot mechanical arm is realized.

5. A neural network-based adaptive trajectory control system for a robot arm for performing the neural network-based adaptive trajectory control method according to any one of claims 1 to 4, characterized in that, The adaptive trajectory control of the mechanical arm through the mechanical arm control torque vector specifically includes: The information acquisition module is configured to collect the state information of the mechanical arm of the music robot in real time, and load the performance task information of the music robot. The bias determination module is configured to generate a performance expected trajectory of each mechanical arm node of the music robot according to the performance task information, and determine a trajectory control bias domain of each mechanical arm node of the music robot in the performance process according to the state information of the mechanical arm and the corresponding performance expected trajectory. The bias compensation module is configured to obtain a pre-trained neural network model, perform bias approximation on the mechanical arm control parameters of the music robot through the neural network model, and obtain trajectory dead zone bias compensation of the mechanical arm control model. The trajectory control module is configured to generate a mechanical arm control torque vector of the music robot according to the trajectory dead zone bias compensation of the mechanical arm and the trajectory control bias domain of each mechanical arm node, and perform adaptive trajectory control of the mechanical arm through the mechanical arm control torque vector.

6. A computer device, comprising: The computer device includes a memory and a processor, the memory stores code, the processor is configured to obtain the code, and execute the adaptive trajectory control method of the mechanical arm based on the neural network in any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, wherein the computer program comprises the following steps of: receiving a request for a resource from a client; determining whether the client is authorized to access the resource; and if the client is authorized to access the resource, providing the resource to the client. The computer program is executed by the processor to realize the adaptive trajectory control method of the mechanical arm based on the neural network in any one of claims 1 to 4.

Images