A dynamic control method for a bionic multi-joint system and related equipment

Abstract

The application discloses a dynamic control method of a bionic multi-joint system and related equipment, and relates to the technical field of automatic control. The method comprises the following steps: acquiring dynamic characteristics of the bionic multi-joint system under different motion tasks in combination with a graph neural network model; establishing a kinematic constraint model for the bionic multi-joint system according to physical boundaries; establishing a dynamics constraint model of a turnover task according to the law of conservation of angular momentum and in combination with the mass distribution of the bionic multi-joint system and a turnover target; solving the kinematic constraint model and the dynamics constraint model based on the dynamic characteristics, the real-time state of the bionic multi-joint system and the turnover task target by using an MPC solver to obtain expected control parameters; and dynamically controlling the bionic multi-joint system according to the expected control parameters. The application can fully utilize the dynamic capability of the bionic multi-joint system and improve the adaptability of the bionic multi-joint system to multiple tasks in combination with the graph neural network model and the MPC solver.

Description

Technical Field

[0001] This application relates to the field of automatic control technology, and in particular to a dynamic control method and related equipment for a biomimetic multi-joint system. Background Technology

[0002] The spine of biological organisms plays a crucial role in enabling complex movements such as bending, extension, lateral flexion, and flipping. These capabilities allow organisms to react quickly to environmental changes and facilitate rapid changes in direction, which is essential for animals to nimbly capture prey or evade predators. Flexible spinal structures enable quadrupedal and humanoid robots to change their posture without external assistance, demonstrating broad application potential in search and rescue, film and television special effects, military reconnaissance, and agriculture. In space exploration, robots with flipping capabilities offer greater flexibility and operational efficiency. These robots can flip and adjust their posture in microgravity environments to assist in maintenance, repair, and construction tasks inside and outside space stations. Despite the critical role of the spine in biological systems, current quadrupedal and humanoid robots often neglect the spine or simplify its complex mechanisms, sacrificing the inherent flexibility of biological systems. This limitation stems primarily from the lack of control algorithms capable of coordinating multi-degree-of-freedom dynamic responses while adapting to diverse task requirements, ultimately restricting the application of the spine in quadrupedal and humanoid robots. Most studies have focused on adjusting the spine in quadruped robots based on predefined models and optimal control algorithms, demonstrating that the spine can improve the walking performance of quadruped robots, such as increasing energy efficiency and stride length. The spine's lateral flexion capability offers unique advantages, including extended stride length, enhanced stability, and a smaller turning radius. However, these works have primarily focused on low-speed tasks and have not explored the dynamic properties of the spine to enhance the robot's flexibility and agility; for example, the spine could assist the robot in performing aerial flips or increasing its speed during movement. Summary of the Invention

[0003] In view of this, embodiments of this application provide a dynamic control method and related equipment for a biomimetic multi-joint system, so as to fully utilize the dynamic capabilities of the biomimetic multi-joint system while improving the adaptability of the biomimetic multi-joint system to multiple tasks.

[0004] One aspect of this application provides a dynamic control method for a biomimetic multi-joint system, the method comprising the following steps:

[0005] By combining graph neural network models, dynamic features of biomimetic multi-joint systems under different motion tasks can be obtained.

[0006] A kinematic constraint model is established for the biomimetic multi-joint system based on the physical boundaries;

[0007] Based on the law of conservation of angular momentum and combined with the mass distribution of the biomimetic multi-joint system and the flipping target, a dynamic constraint model for the flipping task is established.

[0008] The kinematic constraint model and the dynamic constraint model are solved using the MPC solver based on the dynamic characteristics, the real-time state of the biomimetic multi-joint system, and the flipping task objective to obtain the desired control parameters.

[0009] The biomimetic multi-joint system is dynamically controlled according to the desired control parameters.

[0010] In some embodiments, obtaining the dynamic characteristics of the biomimetic multi-joint system under different motion tasks by combining a graph neural network model includes the following steps:

[0011] Obtain a given dataset and a corresponding dynamic parameter set; wherein the given dataset and the dynamic parameter set are both joint information graphs, the joint information graphs are constructed using each joint of the bionic multi-joint system as nodes and the matrix of the connection relationships between each joint as edges, and the point information of the nodes includes the bending angle and rotation angle of the joint.

[0012] The graph neural network model is trained based on the given dataset and the dynamic parameter set, and during training, the difference between the dynamic parameters predicted by the graph neural network model based on the given dataset and the dynamic parameter set is minimized based on the reward function to adjust the parameters of the graph neural network model.

[0013] The joint information graph corresponding to the motion data of the bionic multi-joint system under different motion tasks is input into the trained graph neural network model; the messages of each neighboring node of each node are calculated and aggregated in turn using each layer of the trained graph neural network model; after the last layer, the inertial moment parameter is predicted as the dynamic feature based on the message obtained from the last layer using a multilayer perceptron.

[0014] In some embodiments, the step of sequentially calculating and aggregating the messages of neighboring nodes of each node using each layer of the trained graph neural network model includes the following steps:

[0015] The message is computed at each layer using message computation formulas;

[0016] The message calculation formula is:

[0017] ;

[0018] in, This indicates the message. The message represents the computation of layer k. , This refers to the message aggregated at layer k-1; This indicates the number of layers in the graph neural network model. Indicates the current node, Indicates adjacent nodes. A matrix representing the connection relationships between the various joints. Indicates the bending angle, Indicates the rotation angle; This refers to the multilayer sensor;

[0019] The messages are aggregated at each layer using an aggregation method;

[0020] The polymerization formula is:

[0021] ;

[0022] in, This refers to the message after aggregation at layer k. Represents aggregate functions, Represents a node The set of adjacent nodes.

[0023] In some embodiments, establishing a kinematic constraint model for the biomimetic multi-joint system based on physical boundaries includes the following steps:

[0024] The kinematic constraint model is determined based on the spatial mapping relationship between the end positions of the biomimetic multi-joint system and the state parameters of each joint; wherein, the state parameters include the bending angle and rotation angle of the joint.

[0025] The kinematic constraint model is as follows:

[0026] ;

[0027] ;

[0028] in, Indicates the end position, Let represent the transformation matrix of the i-th key. Indicates the bending angle, Indicates the rotation angle; Represents the initial position vector. and These represent the lower and upper limits of the workspace position at the end of the bionic multi-joint system, respectively.

[0029] In some embodiments, establishing a dynamic constraint model for the flipping task based on the law of conservation of angular momentum and combining the mass distribution of the biomimetic multi-joint system with the flipping target includes the following steps:

[0030] Based on the law of conservation of angular momentum and combined with the mass distribution of the biomimetic multi-joint system and the flipping target, the dynamic constraint model for the flipping task is established as follows:

[0031] ;

[0032] in, This represents the center of mass of the biomimetic multi-joint system. The moment of inertia of the center of mass is represented by the following. Indicates the bending angle of the joint. Indicates the rotation angle of the joint. This represents the overall rotational inertia of the biomimetic multi-joint system. This represents the rotational angular velocity of the biomimetic multi-joint system.

[0033] In some embodiments, the step of using an MPC solver to solve the kinematic constraint model and the dynamic constraint model based on the dynamic characteristics, the real-time state of the biomimetic multi-joint system, and the flipping task objective to obtain the desired control parameters includes the following steps:

[0034] Based on the dynamic characteristics, the real-time state of the biomimetic multi-joint system, and the flipping task objective, the optimization problem for constructing the kinematic constraint model and the dynamic constraint model is as follows:

[0035] ;

[0036] in, This represents the desired control parameter. = { , , This represents the bending angle of the i-th joint in the biomimetic multi-joint system. This represents the rotation angle of the i-th joint in the biomimetic multi-joint system; It is a rotation matrix. This indicates the orientation of the center of mass of the biomimetic multi-joint system; Represent the desired rotation matrix; It is the location of the centroid; Indicates the desired centroid position; , This is the weight matrix; This indicates that the logarithm of the product is taken before the inverse hat operation is performed.

[0037] The constraints of the optimization problem are determined as follows:

[0038] ; ;

[0039] in, and This represents the kinematic constraints of the biomimetic multi-joint system. and This represents the dynamic constraints of the biomimetic multi-joint system; This indicates the system state of the bionic multi-joint system. ;

[0040] The optimization problem is solved using the MPC solver based on the constraints to obtain the desired control parameters.

[0041] In some embodiments, dynamically controlling the bionic multi-joint system according to the desired control parameters includes the following steps:

[0042] The state corresponding to the desired control parameters is converted into the first desired angle and first desired angular velocity of each joint motor in the bionic multi-joint system.

[0043] The trajectory of the bionic multi-joint system is controlled proportionally and derivatively based on the desired angle and the desired angular velocity.

[0044] The expression for the proportional-derivative control is as follows:

[0045] ;

[0046] in, This represents the control torque of the i-th joint. Represents the proportional gain coefficient. Represents the differential gain coefficient. This represents the second desired angle of the i-th joint. This represents the first desired angle. This represents the second desired angular velocity of the i-th joint. This represents the first desired angular velocity.

[0047] Another aspect of this application embodiment provides a biomimetic multi-joint system dynamic control device, the device comprising:

[0048] The feature acquisition unit is used to combine a graph neural network model to acquire the dynamic features of the biomimetic multi-joint system under different motion tasks.

[0049] A kinematic constraint establishment unit is used to establish a kinematic constraint model for the biomimetic multi-joint system based on physical boundaries.

[0050] The dynamic constraint establishment unit is used to establish a dynamic constraint model for the flipping task based on the law of conservation of angular momentum and in combination with the mass distribution of the biomimetic multi-joint system and the flipping target.

[0051] The control parameter solving unit is used to solve the kinematic constraint model and the dynamic constraint model based on the dynamic characteristics, the real-time state of the bionic multi-joint system and the flipping task target using the MPC solver, so as to obtain the desired control parameters.

[0052] A control unit is used to dynamically control the bionic multi-joint system according to the desired control parameters.

[0053] Another aspect of this application embodiment provides an electronic device, including a processor and a memory;

[0054] The memory is used to store programs;

[0055] The processor executes the program to implement any of the methods described above.

[0056] Another aspect of this application provides a computer-readable storage medium storing a program that is executed by a processor to implement the method described in any of the above embodiments.

[0057] This application includes at least the following beneficial effects:

[0058] This application combines a graph neural network model to obtain the dynamic characteristics of a biomimetic multi-joint system under different motion tasks; establishes a kinematic constraint model for the biomimetic multi-joint system based on physical boundaries; establishes a dynamic constraint model for the flipping task based on the law of conservation of angular momentum and the mass distribution of the biomimetic multi-joint system and the flipping target; uses an MPC solver to solve the kinematic and dynamic constraint models based on dynamic characteristics, the real-time state of the biomimetic multi-joint system, and the flipping task target to obtain desired control parameters; and dynamically controls the biomimetic multi-joint system based on the desired control parameters. This application, combining a graph neural network model and an MPC solver, can synergistically optimize the strength, motion accuracy, and compliance of the biomimetic multi-joint system under high load conditions, resolving the technical contradiction of incompatibility among these three aspects in existing technologies. It can fully utilize the dynamic capabilities of the biomimetic multi-joint system while improving its adaptability to multiple tasks. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1A flowchart illustrating a dynamic control method for a biomimetic multi-joint system provided in an embodiment of this application;

[0061] Figure 2 A flowchart illustrating the specific implementation of dynamic control of the biomimetic multi-joint system provided in this application embodiment;

[0062] Figure 3 This is a kinematic diagram of a multi-joint spinal structure provided in an embodiment of this application;

[0063] Figure 4 A schematic diagram showing the spine rotating about a fixed axis with the same curvature angle, as provided in an embodiment of this application;

[0064] Figure 5 This is a diagram showing the rollover trajectory of an active spinal robot during a fall, as provided in an embodiment of this application.

[0065] Figure 6 A schematic diagram of the roll angles of an active spinal robot falling at different initial roll angles, as provided in the embodiments of this application;

[0066] Figure 7 This is a structural block diagram of a biomimetic multi-joint system dynamic control device provided in an embodiment of this application. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0068] Before providing a detailed description of the embodiments of this application, some related technologies involved in the embodiments of this application will be described first, as follows:

[0069] In multi-joint robot research, traditional model-based control methods typically employ predefined models, such as spring-loaded inverted pendulums (SLIPs) or pre-configured system states. These methods improve computational tractability while maintaining adaptability. However, discrepancies between the predefined model and real-world dynamics during motion can lead to system instability. For multi-joint systems, accurate whole-body dynamics are often difficult to obtain, and real-time updates of dynamic parameters are further constrained by computational resources. Therefore, an increasing number of studies are employing learning-based methods to learn accurate system dynamics or combining model-free methods to directly map robot behavior to control policies.

[0070] Current learning-based control methods for multi-joint robots mainly fall into two categories: one involves extensive simulation training in a virtual environment, and the other captures the motion of real organisms for imitation learning. While these methods improve the stability and flexibility of robots, they require large amounts of training data and time, and are typically limited to fixed robot configurations. To overcome these limitations, some research uses machine learning to learn unknown parameters of the dynamics model from limited data. Using a small amount of flight data, deep neural networks (DNNs) can be trained to characterize the aerodynamic effects of quadrotors under different wind conditions. These methods collect prior information about the system, do not require large datasets, are applicable to a wide variety of robot systems, and have the potential to simplify dynamics models.

[0071] While active spines hold promise for enhancing the locomotion performance of legged robots, the complexity of spine models leads many quadruped robots to avoid spinal joints to minimize control complexity. Common controllers for legged robots with active spines are often limited to sagittal plane motion. Recently, there has been increasing interest in compliant spines capable of lateral flexion, which enables legged robots to have smaller turning radii and faster walking speeds. The biomimetic mouse robot NeRmo utilizes a compliant and flexible spine to improve locomotion agility, static stability, walking speed, and maneuverability. However, these studies use open-loop control and trajectory optimization based on simplified models, with the spine trajectory predefined for specific locomotion tasks. To automatically optimize the spine trajectory, some studies employ Q-learning in biomimetic quadruped robots to enhance the dynamic movement of the spine.

[0072] Reference Figure 1 This application provides a dynamic control method for a biomimetic multi-joint system, specifically including the following steps S100~S140:

[0073] S100: Combine graph neural network models to obtain dynamic features of biomimetic multi-joint systems under different motion tasks;

[0074] S110: Establish a kinematic constraint model for the biomimetic multi-joint system based on the physical boundaries;

[0075] S120: Based on the law of conservation of angular momentum and combined with the mass distribution of the biomimetic multi-joint system and the flipping target, establish a dynamic constraint model for the flipping task;

[0076] S130: Using the MPC solver, the kinematic constraint model and the dynamic constraint model are solved based on the dynamic characteristics, the real-time state of the bionic multi-joint system, and the flipping task target to obtain the desired control parameters;

[0077] S140: Dynamically control the bionic multi-joint system according to the desired control parameters.

[0078] Optionally, the step of obtaining the dynamic characteristics of the bionic multi-joint system under different motion tasks by combining a graph neural network model includes the following steps:

[0079] Obtain a given dataset and a corresponding dynamic parameter set; wherein the given dataset and the dynamic parameter set are both joint information graphs, the joint information graphs are constructed using each joint of the bionic multi-joint system as nodes and the matrix of the connection relationships between each joint as edges, and the point information of the nodes includes the bending angle and rotation angle of the joint.

[0080] The graph neural network model is trained based on the given dataset and the dynamic parameter set, and during training, the difference between the dynamic parameters predicted by the graph neural network model based on the given dataset and the dynamic parameter set is minimized based on the reward function to adjust the parameters of the graph neural network model.

[0081] The joint information graph corresponding to the motion data of the bionic multi-joint system under different motion tasks is input into the trained graph neural network model; the messages of each neighboring node of each node are calculated and aggregated in turn using each layer of the trained graph neural network model; after the last layer, the inertial moment parameter is predicted as the dynamic feature based on the message obtained from the last layer using a multilayer perceptron.

[0082] Optionally, the step of sequentially calculating and aggregating the messages of each neighboring node using each layer of the trained graph neural network model includes the following steps:

[0083] The message is computed at each layer using message computation formulas;

[0084] The message calculation formula is:

[0085] ;

[0086] in, This indicates the message. The message represents the computation of layer k. , This refers to the message aggregated at layer k-1; This indicates the number of layers in the graph neural network model. Indicates the current node, Indicates adjacent nodes. A matrix representing the connection relationships between the various joints. Indicates the bending angle, Indicates the rotation angle; This refers to the multilayer sensor;

[0087] The messages are aggregated at each layer using an aggregation method;

[0088] The polymerization formula is:

[0089] ;

[0090] in, This refers to the message after aggregation at layer k. Represents aggregate functions, Represents a node The set of adjacent nodes.

[0091] Optionally, establishing a kinematic constraint model for the biomimetic multi-joint system based on physical boundaries includes the following steps:

[0092] The kinematic constraint model is determined based on the spatial mapping relationship between the end positions of the biomimetic multi-joint system and the state parameters of each joint; wherein, the state parameters include the bending angle and rotation angle of the joint.

[0093] The kinematic constraint model is as follows:

[0094] ;

[0095] ;

[0096] in, Indicates the end position, Let represent the transformation matrix of the i-th key. Indicates the bending angle, Indicates the rotation angle; Represents the initial position vector. and These represent the lower and upper limits of the workspace position at the end of the bionic multi-joint system, respectively.

[0097] Optionally, the step of establishing a dynamic constraint model for the flipping task based on the law of conservation of angular momentum and combining the mass distribution of the biomimetic multi-joint system with the flipping target includes the following steps:

[0098] Based on the law of conservation of angular momentum and combined with the mass distribution of the biomimetic multi-joint system and the flipping target, the dynamic constraint model for the flipping task is established as follows:

[0099] ;

[0100] in, This represents the center of mass of the biomimetic multi-joint system. The moment of inertia of the center of mass is represented by the following. Indicates the bending angle of the joint. Indicates the rotation angle of the joint. This represents the overall rotational inertia of the biomimetic multi-joint system. This represents the rotational angular velocity of the biomimetic multi-joint system.

[0101] Optionally, the step of using the MPC solver to solve the kinematic constraint model and the dynamic constraint model based on the dynamic characteristics, the real-time state of the biomimetic multi-joint system, and the flipping task objective to obtain the desired control parameters includes the following steps:

[0102] Based on the dynamic characteristics, the real-time state of the biomimetic multi-joint system, and the flipping task objective, the optimization problem for constructing the kinematic constraint model and the dynamic constraint model is as follows:

[0103] ;

[0104] in, This represents the desired control parameter. = { , , This represents the bending angle of the i-th joint in the biomimetic multi-joint system. This represents the rotation angle of the i-th joint in the biomimetic multi-joint system; It is a rotation matrix. This indicates the orientation of the center of mass of the biomimetic multi-joint system; Represent the desired rotation matrix; It is the location of the centroid; Indicates the desired centroid position; , This is the weight matrix; This indicates that the logarithm of the product is taken before the inverse hat operation is performed.

[0105] The constraints of the optimization problem are determined as follows:

[0106] ; ;

[0107] in, and This represents the kinematic constraints of the biomimetic multi-joint system. and This represents the dynamic constraints of the biomimetic multi-joint system; This indicates the system state of the bionic multi-joint system. ;

[0108] The optimization problem is solved using the MPC solver based on the constraints to obtain the desired control parameters.

[0109] Optionally, dynamically controlling the bionic multi-joint system according to the desired control parameters includes the following steps:

[0110] The state corresponding to the desired control parameters is converted into the first desired angle and first desired angular velocity of each joint motor in the bionic multi-joint system.

[0111] The trajectory of the bionic multi-joint system is controlled proportionally and derivatively based on the desired angle and the desired angular velocity.

[0112] The expression for the proportional-derivative control is as follows:

[0113] ;

[0114] in, This represents the control torque of the i-th joint. Represents the proportional gain coefficient. Represents the differential gain coefficient. This represents the second desired angle of the i-th joint. This represents the first desired angle. This represents the second desired angular velocity of the i-th joint. This represents the first desired angular velocity.

[0115] The following section will provide a detailed introduction and explanation of the solutions in the embodiments of this application, using specific application examples.

[0116] This embodiment can be applied to the field of biomimetic robot control technology, specifically involving a dynamic and precise control method for multi-joint continuous systems. Addressing the problems of insufficient flexibility, limited load capacity, and poor real-time performance in the dynamic control of existing biomimetic multi-joint systems (such as biomimetic spinal systems), this embodiment proposes a general control scheme integrating online and offline learning. This scheme uses a graph neural network (GNN) to learn the parameters of the spinal dynamics model offline, constructing a high-precision nonlinear dynamic model; simultaneously, it combines parametric model predictive control (MPC) to update dynamic constraints online and select the optimal control strategy in real time. In a spinal aerial flipping task, by analyzing the angular momentum dynamic constraints of the multi-joint structure, an efficient flipping strategy that does not rely on external forces is generated, enabling the spinal system to autonomously and continuously flip in the air. Further experiments show that this embodiment can synergistically optimize the strength, motion accuracy, and compliance of the spine under high load conditions, resolving the technical contradiction of incompatibility among these three aspects in traditional methods. Experiments show that the spinal system using the GNN-MPC controller can simulate the flipping behavior of a biological spine in an aerial flipping task. This embodiment can be widely applied in fields such as biomimetic robots and high-load flexible robotic arms.

[0117] This embodiment can be applied to the field of flexible control technology for bionic robots, specifically involving a whole-body planning and control strategy for bionic robots oriented towards flexible tasks. It is particularly suitable for complex tasks requiring quadrupedal and humanoid robots, such as inspection and disaster relief tasks in complex environments.

[0118] Specifically, this embodiment may include the following steps:

[0119] This embodiment relates to a learning-based active spine control method to improve the motion flexibility, stability, and environmental adaptability of robots equipped with an active spine in complex scenarios. This method integrates graph neural networks (GNNs) and parametric model predictive control (MPC) (GNN-MPC) to achieve precise understanding and control of the spine's dynamic characteristics, thereby optimizing the robot's overall motion performance. Specifically, this embodiment includes the following key innovative steps:

[0120] Step 1: Through offline data acquisition and graph neural network algorithms, the dynamic constraint characteristics and potential motion patterns of the active spine under various working conditions are learned for different robot motion tasks (such as ground walking and aerial flipping). A large amount of spinal motion data is analyzed using graph neural networks (GNNs) to establish an initial dynamic feature model of the spine, providing dynamic parameter support for subsequent control strategies.

[0121] Step Two: Conduct kinematic analysis of the spine, focusing on the kinematic constraints of the spinal joints. Through precise measurement and modeling, determine the angular limits, range of motion, and inter-joint linkage limitations of each spinal joint in flexion, extension, and torsion directions, constructing an accurate kinematic constraint model. This model defines the physical boundaries of spinal movement, avoiding mechanical damage caused by exceeding movement limits, and provides clear kinematic boundary conditions for subsequent control strategies, ensuring that spinal movement conforms to mechanical characteristics.

[0122] Step 3: Analyze the dynamic constraints of the spine for the flipping task. Conduct an in-depth study of the law of conservation of angular momentum followed by the spine during the flipping process. Combine the robot's mass distribution with the flipping objectives (such as flipping speed and angular accuracy) to establish a spine dynamic constraint model specifically for the flipping task, ensuring that the flipping process meets both dynamic requirements and maintains structural safety.

[0123] Step 4: Solve using the MPC solver. Input the kinematic constraint model from Step 2 and the dynamic constraint model from Step 3 into the MPC solver. Combining the robot's real-time state (e.g., current joint angles and angular velocities) with the flipping task objective, the solver uses optimization algorithms (e.g., quadratic programming) to solve for the optimal control inputs (e.g., motor drive torque, motion commands) for each joint of the spine within each control cycle. Based on the constraints and prediction model, the MPC solver proactively plans the spinal motion trajectory, achieving accurate prediction and dynamic adjustment of the flipping motion, ensuring the flipping task is completed efficiently, stably, and accurately.

[0124] Step 5: Repeat steps 2 through 4 until the task is completed.

[0125] The following describes a more specific implementation method.

[0126] The following example, using the adaptive control of a bionic multi-joint system carrying an active spine in performing an aerial flipping task, will explain the technical solution of this embodiment in more detail.

[0127] Reference Figure 2 This embodiment provides a flowchart of the specific implementation of dynamic control of a biomimetic multi-joint system.

[0128] Step 1: Analyze the dynamic parameter characteristics of the robot's spine and introduce a model parameter recognition method based on GNN.

[0129] When a robot performs flips and runs, the corresponding dynamic constraints of the spine differ. Accurate dynamic parameters are crucial for analyzing the dynamic model of the spine under various conditions. The dynamic parameters of the robot's spine (such as the inertia matrix) change with the robot's joint states; therefore, it is necessary to identify the relationship between dynamic parameters and robot joint states. In multi-joint cascaded active spine structures, although the transmission method is highly efficient, there is a problem of uneven mass distribution, making it extremely difficult to determine the intrinsic relationship between dynamic parameters and joint angles using traditional dynamic analysis methods (such as the Lagrangian method).

[0130] The Body Transformer architecture based on Generative Neural Networks (GNNs) is used to learn robot policies. It can capture the structural information and relationships of multi-joint robots, performs well in multiple learning tasks, and has high computational efficiency, making it suitable for multi-joint robot learning. Therefore, a model parameter recognition method based on GNNs is adopted. The GNN input consists of point information related to robot joints and edge information representing the connection relationships between joints.

[0131] like Figure 3 As shown, the point information includes the spinal joint curvature angle. and rotation angle Let each joint be represented as a node in the graph. For a spinal robot with n joints, there is a set of points. Each point With the Feature vectors of each joint Related. Edge information is a matrix representing the connection relationships between joints. If the joint and Connect, then ;otherwise, .

[0132] A GNN consists of multiple layers, each performing message passing and feature aggregation. In the first layer, for each node... Using formula Calculate from neighboring nodes The message, among which and These are nodes and initial characteristics, It is a multilayer perceptron (MLP), whose input includes node features, edge connectivity information, and differences in joint bending and rotation angles between adjacent nodes, in order to capture local relationships between joints. After computing the messages, an aggregation function is used. (Maximum function) for each node polymerization: , yes The set of adjacent nodes.

[0133] This process is repeated for multiple layers. In subsequent layers k, the message calculation formula is: The aggregation formula is Finally, on the last floor Then, the moment of inertia parameters were predicted using MLP. : = This GNN architecture aims to capture the complex relationships between joint information in a spinal robot and effectively predict dynamic parameters.

[0134] Given dataset and corresponding dynamic parameter set The goal of the reward function is to encourage GNNs to perform well. Given a spinal state Accurately predict dynamic parameters. Define the reward function. as follows: ,in GNN as input The corresponding predictive dynamic parameters, It comes from The true dynamic parameters are then used. The negative sign transforms the error into the reward to be maximized. By minimizing the difference between the predicted and true values ​​of the moment of inertia, the GNN is trained to better approximate the relationship between the input configuration and the dynamic parameters.

[0135] Step Two: As before Figure 3 As shown, the spinal structure is modeled as a series-compatible rotational joint system with n degrees of freedom.

[0136] Based on the physical constraints of spinal joints, the rotation angles of each rotational joint must satisfy consistency constraints. The transformation matrix of the i-th joint module is defined as follows: ,in For rotation angle, Let be the bending angle. A kinematic constraint model is established for the spinal system, which consists of j connected joint modules. This is achieved by analyzing the position of the spinal end. With the state parameters of each joint Based on the spatial mapping relationship, the relative position constraint equations are derived as follows:

[0137] ;

[0138] ;

[0139] in, Initial position vector, and Let represent the lower and upper limits of the workspace at the end of the spine, respectively. The above constraint equations show that as the number of tandem joints j increases, the workspace range at the end of the spine expands, but the system's load capacity and motion accuracy will decrease inversely, requiring balance optimization in actual control.

[0140] Step 3: For the dynamic characteristic of the spinal robot that it can flip over using only the spine, construct dynamic constraints based on the law of conservation of angular momentum.

[0141] When a robot falls through the air with zero initial angular momentum, its overall angular momentum must be conserved. Assuming the robot is only subject to downward gravity and its center of mass moves along the z-axis of the world coordinate system, we can analyze the spinal curvature angle... With rotation angle The effect on the system's angular momentum divides the spine into three parts, each about an axis of rotation passing through the center of mass. The generated angular momentum are respectively , , Its vector sum is + + According to the law of conservation of angular momentum, there must exist another angular momentum of equal magnitude and opposite direction in the robot system. To offset This keeps the total angular momentum constant. Generated by the robot's own rotation around its axis, satisfying: .

[0142] Since angular momentum equals the product of moment of inertia and angular velocity, the entire spine rotates around... Moment of inertia It is the sum of the moments of inertia of each part. And the angular velocities of each part are equal. Therefore:

[0143] ;

[0144] in, For the robot's overall rotational inertia, Let be the robot's own rotational angular velocity. Combining the law of conservation of energy, when... At this time, the spinal rotational kinetic energy is completely converted into the robot's rotational kinetic energy, resulting in the highest flipping efficiency and no energy loss; if much smaller The spinal rotational angular velocity is much greater than the robot's rotational angular velocity, resulting in significant kinetic energy loss and low flipping efficiency. Therefore, equation "..." is chosen. "This serves as a dynamic constraint for the robot's mid-air flips. The solver considers the moment of inertia trained in step one." With spinal condition Based on this relationship, the control output with the highest flipping efficiency can be calculated, ensuring efficient attitude adjustment while satisfying the conservation of angular momentum. For example, Figure 4 This is a schematic diagram of the spine rotating about a fixed axis with the same curvature angle. With the same value, changing only α will cause the spine to rotate about a fixed axis with the same curvature angle, thus generating rotational angular momentum.

[0145] Step 4: Solving the MPC optimization problem.

[0146] This embodiment utilizes the curvature of the spine. and rotation angle This describes the rotational state of the spinal joints. The system state of the spinal robot is determined by... Here, It is a rotation matrix, representing the orientation of the robot's center of mass. It is the location of the robot's center of mass. and This represents the state of the i-th vertebral joint, and the control output is selected as... = { , ,in The optimization problem is to achieve the desired state. As input, the output is the trajectory of the spine. Therefore, this embodiment constructs the following optimization problem:

[0147] ;

[0148] The goal is to minimize an expression consisting of two parts. The first part is the error concerning the centroid position, i.e., the centroid position... With respect to the expected centroid position The weighted square norm of the transpose difference, the weight matrix is The second part concerns the error in the centroid's attitude, specifically the rotation matrix. With the desired rotation matrix The transpose of the product is then logarithmed and then the inverse hash operation is performed (here, using...). The weighted square norm of (represented by) is given by the weight matrix as follows: .

[0149] At the same time, the optimization problem needs to satisfy the following constraints:

[0150] ; ;

[0151] in, and The kinematic constraints represent those of the spinal robot in step two. and This represents the dynamic constraints of the spinal robot in step three.

[0152] Because the spine functions differently in different scenarios, the corresponding dynamic constraints also differ. To improve computational efficiency, the control system in this embodiment analyzes the dynamic model of the spine according to different scenarios and selects the most effective dynamic constraints. Meanwhile, the kinematic constraints of the same vertebral joint remain unchanged.

[0153] By specifying dynamic constraints for different tasks and solving this optimization problem, this embodiment can obtain the desired output of the spinal system. Then, through inverse kinematics calculations, The corresponding state is converted into the desired angle of each joint motor. and expected angular velocity Joint trajectories are achieved through proportional-derivative (PD) control:

[0154] ;

[0155] in, It is the control torque of the i-th joint. It is the proportional gain coefficient. It is the differential gain coefficient. It is the expected angle of the i-th joint. This is the expected angular velocity of the i-th joint. In this way, this embodiment can achieve effective control of the joints of the spinal robot, enabling it to move flexibly according to different task requirements and scenarios.

[0156] Step 5: Repeat steps 3 and 4 until the task is completed.

[0157] For example, in the experiment of this embodiment, Figure 5 A diagram showing the flipping motion trajectory of an active spinal robot falling from a height of 2.8 meters. Figure 6 A schematic diagram showing the roll angles of an active spinal robot falling at different initial roll angles.

[0158] In summary, this embodiment includes the following technical solutions and beneficial effects:

[0159] 1. Dynamic parameter identification method based on GNN: To address the problem of uneven mass distribution in multi-joint series structures and the difficulty of modeling by traditional dynamic analysis methods, a graph neural network (GNN) is introduced to capture the complex relationships between joints. By using node features and edge connection information, the dynamic parameters of the spine can be efficiently predicted, solving the modeling problem of dynamic parameters changing with joint state.

[0160] 2. Dynamic constraint model of spinal rollover based on conservation of angular momentum: Establish a quantitative relationship between the bending / rotation angle of the spine and the system's moment of inertia and angular velocity during aerial rollover, ensuring conservation of angular momentum and efficient energy conversion during the rollover process, and achieving autonomous attitude adjustment without external torque.

[0161] 3. Scene-adaptive dynamic constraint selection mechanism: To address the differences in spinal function in different tasks (such as flipping and walking), a dynamic constraint adaptive selection algorithm is designed. The optimal dynamic model is switched according to the real-time scene, which improves the robustness and flexibility of the system in complex environments while keeping the kinematic constraints unchanged.

[0162] Reference Figure 7 This application provides a biomimetic multi-joint system dynamic control device, comprising:

[0163] The feature acquisition unit is used to combine a graph neural network model to acquire the dynamic features of the biomimetic multi-joint system under different motion tasks.

[0164] A kinematic constraint establishment unit is used to establish a kinematic constraint model for the biomimetic multi-joint system based on physical boundaries.

[0165] The dynamic constraint establishment unit is used to establish a dynamic constraint model for the flipping task based on the law of conservation of angular momentum and in combination with the mass distribution of the biomimetic multi-joint system and the flipping target.

[0166] The control parameter solving unit is used to solve the kinematic constraint model and the dynamic constraint model based on the dynamic characteristics, the real-time state of the bionic multi-joint system and the flipping task target using the MPC solver, so as to obtain the desired control parameters.

[0167] A control unit is used to dynamically control the bionic multi-joint system according to the desired control parameters.

[0168] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0169] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0170] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0171] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0172] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0173] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0174] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0175] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0176] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0177] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A dynamic control method for a biomimetic multi-joint system, characterized in that, The method includes the following steps: By combining graph neural network models, dynamic features of biomimetic multi-joint systems under different motion tasks can be obtained. A kinematic constraint model is established for the biomimetic multi-joint system based on the physical boundaries; Based on the law of conservation of angular momentum and combined with the mass distribution of the biomimetic multi-joint system and the flipping target, a dynamic constraint model for the flipping task is established. The kinematic constraint model and the dynamic constraint model are solved using the MPC solver based on the dynamic characteristics, the real-time state of the biomimetic multi-joint system, and the flipping task objective to obtain the desired control parameters. The bionic multi-joint system is dynamically controlled according to the desired control parameters. The process of obtaining the dynamic characteristics of the biomimetic multi-joint system under different motion tasks by combining a graph neural network model includes the following steps: Obtain a given dataset and a corresponding dynamic parameter set; wherein the given dataset and the dynamic parameter set are both joint information graphs, the joint information graphs are constructed using each joint of the bionic multi-joint system as nodes and the matrix of the connection relationships between each joint as edges, and the point information of the nodes includes the bending angle and rotation angle of the joint. The graph neural network model is trained based on the given dataset and the dynamic parameter set, and the graph neural network model is minimized based on the reward function during training. The parameters of the graph neural network model are adjusted based on the difference between the dynamic parameters predicted by the given dataset and the dynamic parameter set. The joint information graph corresponding to the motion data of the bionic multi-joint system under different motion tasks is input into the trained graph neural network model; the messages of each neighboring node of each node are calculated and aggregated in turn using each layer of the trained graph neural network model; after the last layer, the inertial moment parameter is predicted as the dynamic feature based on the message obtained from the last layer using a multilayer perceptron. The step of sequentially calculating and aggregating the messages of adjacent nodes of each node using each layer of the trained graph neural network model includes the following steps: The message is computed at each layer using message computation formulas; The message calculation formula is: ； in, This indicates the message. The message represents the computation of layer k. , This refers to the message aggregated at layer k-1; This indicates the number of layers in the graph neural network model. Indicates the current node, Indicates adjacent nodes, A matrix representing the connection relationships between the various joints. Indicates the bending angle, Indicates the rotation angle; This refers to the multilayer sensor; The messages are aggregated at each layer using an aggregation method; The aggregation formula is: ； in, This refers to the message after aggregation at layer k. Represents aggregate functions, Represents a node The set of adjacent nodes; The process of using an MPC solver to solve the kinematic constraint model and the dynamic constraint model based on the dynamic characteristics, the real-time state of the biomimetic multi-joint system, and the flipping task objective to obtain the desired control parameters includes the following steps: Based on the dynamic characteristics, the real-time state of the biomimetic multi-joint system, and the flipping task objective, the optimization problem for constructing the kinematic constraint model and the dynamic constraint model is as follows: ； in, This represents the desired control parameter. = { , , This represents the bending angle of the i-th joint in the biomimetic multi-joint system. This represents the rotation angle of the i-th joint in the biomimetic multi-joint system; It is a rotation matrix. This indicates the orientation of the center of mass of the biomimetic multi-joint system; Represent the desired rotation matrix; It is the location of the centroid; Indicates the desired centroid position; , This is the weight matrix; This indicates that the logarithm of the product is taken before the inverse hat operation is performed. The constraints of the optimization problem are determined as follows: ； ； in, and This represents the kinematic constraints of the biomimetic multi-joint system. and This represents the dynamic constraints of the biomimetic multi-joint system; This indicates the system state of the bionic multi-joint system. ; The optimization problem is solved using the MPC solver based on the constraints to obtain the desired control parameters.

2. The dynamic control method for a biomimetic multi-joint system according to claim 1, characterized in that, The process of establishing a kinematic constraint model for the biomimetic multi-joint system based on physical boundaries includes the following steps: The kinematic constraint model is determined based on the spatial mapping relationship between the end positions of the biomimetic multi-joint system and the state parameters of each joint; wherein, the state parameters include the bending angle and rotation angle of the joint. The kinematic constraint model is as follows: ； ； in, Indicates the end position, Let represent the transformation matrix of the i-th key. Indicates the bending angle, Indicates the rotation angle; Represents the initial position vector. and These represent the lower and upper limits of the workspace position at the end of the bionic multi-joint system, respectively.

3. The dynamic control method for a biomimetic multi-joint system according to claim 1, characterized in that, The process of establishing a dynamic constraint model for the flipping task based on the law of conservation of angular momentum and combining the mass distribution of the biomimetic multi-joint system with the flipping target includes the following steps: Based on the law of conservation of angular momentum and combined with the mass distribution of the biomimetic multi-joint system and the flipping target, the dynamic constraint model for the flipping task is established as follows: ； in, This represents the center of mass of the biomimetic multi-joint system. The moment of inertia of the center of mass is represented by the following. Indicates the bending angle of the joint. Indicates the rotation angle of the joint. This represents the overall rotational inertia of the biomimetic multi-joint system. This represents the rotational angular velocity of the biomimetic multi-joint system.

4. A dynamic control method for a biomimetic multi-joint system according to any one of claims 1 to 3, characterized in that, The dynamic control of the bionic multi-joint system according to the desired control parameters includes the following steps: The state corresponding to the desired control parameters is converted into the first desired angle and first desired angular velocity of each joint motor in the bionic multi-joint system. The trajectory of the bionic multi-joint system is controlled proportionally and derivatively based on the desired angle and the desired angular velocity. The expression for the proportional-derivative control is as follows: ； in, This represents the control torque of the i-th joint. Represents the proportional gain coefficient. Represents the differential gain coefficient. This represents the second desired angle of the i-th joint. Indicates the first desired angle. This represents the second desired angular velocity of the i-th joint. This represents the first desired angular velocity.

5. A biomimetic multi-joint system dynamic control device, characterized in that, The device is used to implement the dynamic control method for a biomimetic multi-joint system as described in claim 1, and the device includes: The feature acquisition unit is used to combine a graph neural network model to acquire the dynamic features of the biomimetic multi-joint system under different motion tasks. A kinematic constraint establishment unit is used to establish a kinematic constraint model for the biomimetic multi-joint system based on physical boundaries. The dynamic constraint establishment unit is used to establish a dynamic constraint model for the flipping task based on the law of conservation of angular momentum and in combination with the mass distribution of the biomimetic multi-joint system and the flipping target. The control parameter solving unit is used to solve the kinematic constraint model and the dynamic constraint model based on the dynamic characteristics, the real-time state of the bionic multi-joint system and the flipping task target using the MPC solver, so as to obtain the desired control parameters. A control unit is used to dynamically control the bionic multi-joint system according to the desired control parameters.

6. An electronic device, characterized in that, The electronic device includes a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 4.

Images