A humanoid robot trunk motion planning method

Abstract

A method for motion planning of a humanoid robot trunk belongs to the technical field of humanoid robots. The method comprises using human skeleton point recognition technology to recognize human demonstration movements; constructing a skeleton recognition point set; establishing a mapping relationship from the Cartesian space of the skeleton recognition point set to the joint space of the robot trunk, and obtaining a trunk joint angle data set of the robot; using a neural network to train the relationship between the coordinate position of the skeleton recognition point and the trunk joint angle data based on imitation learning; using the output obtained by the neural network training as the input of a second-order system model, and calculating the desired trunk joint angle and angular velocity through the second-order system model to complete the motion planning of the humanoid robot trunk. The present invention is applied to a humanoid robot, has the advantages of energy saving and flexible operation, and realizes the motion planning effect of the trunk adaptive double-arm position.

Description

Technical Field

[0001] The invention relates to a robot trunk motion control method, in particular to a humanoid robot trunk motion planning method, and belongs to the technical field of humanoid robots. Background Art

[0002] Humanoid robots are becoming more and more human-like in configuration design. The torso and lower limb structures expand the operating range of the arms. Currently, the focus of humanoid robot research is on the motion planning of the arms, and there is little research on the torso movement of humanoid robots.

[0003] In terms of mechanical configuration, there have been many humanoid designs of robots, but current research is still focused on the upper and lower limbs. Since complex trunk mechanical structure design will increase the difficulty of whole-body control, a simplified approach is often used for the trunk part. The trunk design of most robots is a box-shaped body with only 0-3 degrees of freedom. In order to make the humanoid structure of the robot have the excellent performance of the human trunk, some research has been conducted on the structural design of the robot trunk, such as the flexible spine humanoid robot, which has 5 ball joints in the spine. In order to achieve the balance of the humanoid bipedal robot, the movement of the trunk is calculated in the article "Development of a bipedal walking robot with three-axis torque compensation through trunk motion" to compensate for the torque generated by the movement of the lower and upper limbs.

[0004] For robots with lower limb motion mechanisms, the movement of the torso needs to consider the overall balance of the movement. However, when the lower limbs are fixed, the focus is on the flexibility and comfort of the robot's movements. The article "From the Human Spine to the Humanoid Robot Trunk" divides the movement of the human trunk into three modes of movement, including forward and backward flexion and extension, rotation, and lateral bending through the study of the human spine, muscles, and ligaments. Based on the three modes of movement of the trunk, and taking into account the complexity of the structural design, the difficulty of control, and the movement effect, we designed a torso with a three-degree-of-freedom full-drive series structure in the humanoid upper limb robot. However, how to make the robot's torso move like a human torso remains to be studied.

[0005] There is still a lack of work on individual motion planning for the trunk of a humanoid upper limb robot. The traditional pseudo-inverse or damped least squares method can simultaneously solve the angular velocity of each joint of a humanoid upper limb robot, but it does not divide the functions of different joints and cannot ensure the natural movements of the humanoid robot. Some literature creates a capability graph based on the accessibility index. In order to reach the target with more options, the position of the robot's trunk should make the target located in the area between 68 and 76 in the accessibility index of the capability graph. On this basis, combined with the capability graph, the movement position of the mobile chassis during the cabinet opening process can be determined. However, the configuration of the trunk remains fixed during the execution of the dual-arm task, and the more accurate the capability graph is, the larger the storage space required. In addition, the article "A whole-body control framework with speed-level decoupling" mentioned that the robot's trunk is regarded as a floating base, which can be decoupled from the arms and head. Based on this control framework, the trunk can maintain a certain configuration unchanged during the dual-arm operation, or it can move with the movement of the arms.

[0006] Learning directly from human behavior can eliminate the complex modeling process in the robot motion planning process. For redundant robot arms, optimization goals can be established to achieve the desired motion. However, for humanoid robots with high redundancy, it is difficult to establish an accurate optimization objective function. Therefore, using imitation learning can avoid the modeling process and directly transfer human optimization actions to the robot. However, there are three difficulties in using imitation learning. The first difficulty is how to collect expert data for learning. Common methods for collecting expert data include drag teaching, wearable motion capture systems, and visual cameras; the second difficulty is how to apply expert data to robots; the configuration differences between the human body and the robot make the collected expert data unable to be directly applied to the robot; the third difficulty is how to determine the characteristics of the network input; the robot has a complex structure and many degrees of freedom, resulting in the system having more feature information, and input without pre-processing will greatly reduce the effect of network learning.

[0007] In view of the above, it is urgent to study how to generate humanoid torso movements during the collaborative operation of humanoid upper limb robots to achieve better arm movements. Summary of the invention

[0008] In order to overcome the defects of the prior art, the present invention further provides a humanoid robot trunk motion planning method. The method is used to achieve the change of the trunk joint angle so that the ends of the two arms maintain a high degree of maneuverability, improve the flexibility of the robot, facilitate the task operation of the two arms, and minimize the overall energy consumption of the robot.

[0009] A humanoid robot trunk motion planning method comprises the following steps:

[0010] S1. Use human skeleton recognition technology to identify human demonstration movements;

[0011] S2, construct a skeleton recognition point set;

[0012] S3, establishing a mapping relationship from the Cartesian space of the skeleton recognition point set to the joint space of the robot trunk, and obtaining a trunk joint angle data set of the robot;

[0013] S4, using a neural network to train the relationship between the coordinate positions of the skeleton recognition points and the trunk joint angle data based on imitation learning;

[0014] S5. The output obtained from the neural network training is used as the input of the second-order system model. The desired trunk joint angle and angular velocity are calculated by the second-order system model to complete the motion planning of the humanoid robot trunk.

[0015] Furthermore, in step S1, a ZED camera is used to recognize human demonstration movements.

[0016] Furthermore, the skeleton recognition point set of step S2 includes skeleton recognition points of a human body demonstration and skeleton recognition points obtained by randomly extracting a data set from the skeleton recognition points of a human body demonstration using a data enhancement method.

[0017] Furthermore, the specific process of the data enhancement method for obtaining skeleton identification points is: randomly extract two points as a group from the existing human body demonstration skeleton point data, perform linear interpolation, and obtain one-quarter, one-half and three-quarter points each time to obtain three groups of skeleton identification points, repeat the random extraction times and interpolation, and finally obtain several groups of skeleton identification points after data enhancement.

[0018] Furthermore, the mapping relationship between the Cartesian space of the skeleton identification point set and the joint space of the robot trunk is established as follows:

[0019] The ZED camera recognizes the skeleton points of the upper body, and uses skeleton points 0, 3, 4, and 11 as key points to establish the human body coordinate system as X B Y B Z B , the coordinate origin O coincides with the skeleton point 0, X B The direction from skeleton point 0 to skeleton point 22, Z B is the vertical upward direction, Y B Determined by the right hand system; the angle value of the first torso joint θ t1 Defined as the line connecting skeleton points 0 and 3 and Z B The angle between the directions. Define the plane perpendicular to the line between 0 and 3 as P1, and the angle value of the second torso joint θ t2 Defined as the projection of the line connecting skeleton points 4 and 11 on P1 and X B The angle of the third trunk joint is θt3 Defined as the angle between the line connecting skeleton points 4 and 11 and the projection of this line on P1.

[0020] In step S4, a fully connected neural network is used to learn the relationship between the coordinate positions of skeleton points 7 and 14 and the trunk joint angle data in step S3.

[0021] Furthermore, the second-order system model in step S5 is expressed as follows:

[0022]

[0023] In the formula, the subscript i represents the i-th joint of the trunk, M i represents the system quality of the second-order system, K di represents the damping coefficient of the second-order system, K si represents the stiffness coefficient of the second-order system, θ ti , and represent the desired trunk joint angle, angular velocity and angular acceleration respectively, and F ei Represents the system external force input of the second-order system, which is equal to the output value of the neural network training. The expected joint angle and angular velocity after processing by the second-order system are obtained through the above formula.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] The present invention mainly studies how to generate humanoid trunk movements in the collaborative operation process of humanoid upper limb robots to achieve better arm movements. In view of the above-mentioned difficulty analysis, firstly, an inverse kinematics framework that decouples the trunk from the arms and the head is proposed, and based on this framework, a cascade control logic of "target object-arms-torso" is creatively proposed. Different from directly incorporating the trunk degrees of freedom into the visual motion planning of the head, we use the end positions of the arms to drive the movement of the torso from the perspective of improving the flexibility of the arms. Then, imitation learning is applied to the motion planning of the humanoid robot torso, and a visual camera is used to obtain the human task operation information. In the process of human skeleton recognition points in the robot's robot operation, it is innovatively proposed to apply the data enhancement method to speed up the sampling of human estimated recognition points, and establish a mapping relationship from the Cartesian space of human skeleton recognition points to the joint space of the robot torso, thereby obtaining an expert data set for neural network training; secondly, it is innovatively proposed to use the second-order system model to realize the hybrid architecture of the torso neural network and the velocity-level inverse kinematics of the two arms; finally, the torso motion planning method proposed by this method is compared with the damped least squares method and the torso fixed method in terms of the maneuverability of the two arms and the total energy consumption of the robot, verifying the feasibility and advantages of this method.

[0026] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flow chart of the humanoid robot trunk motion planning method of the present invention;

[0028] Figure 2 This is a configuration diagram of a humanoid upper limb robot;

[0029] Figure 3 It is a step implementation diagram of the humanoid robot trunk motion planning method in the embodiment;

[0030] Figure 4 This is a rendering of the data collection and data mapping in the embodiment;

[0031] Figure 5 Left and right wrist trajectory point diagrams collected for human body demonstration in the embodiment;

[0032] Figure 6 The left and right wrist trajectory point diagrams obtained by data enhancement in the embodiment;

[0033] Figure 7 A geometric mapping relationship diagram from Cartesian space to joint space in an embodiment;

[0034] Figure 8 A comparison diagram of arm-lifting actions of three methods in the embodiment;

[0035] FIG9( a ) is a comparison diagram of the angles and angular velocities of the three joints of the trunk in the trunk motion planning method of the present application in an embodiment;

[0036] FIG9( b ) is a comparison diagram of the angles and angular velocities of the three joints of the trunk in the traditional damped least squares method in the embodiment;

[0037] FIG9( c ) is a comparison diagram of the angles and angular velocities of the three joints of the trunk in the conventional method of fixing the trunk in the embodiment;

[0038] FIG10( a ) is a graph showing the change in the maneuverability of the left arm over time in the three methods of the embodiment;

[0039] FIG10( b ) is a graph showing the change in the operability of the right arm over time in the three methods of the embodiment;

[0040] FIG10( c ) is a comparison diagram of the energy consumption of the robot over time in the three methods of the embodiment;

[0041] Fig.11 This is a process diagram of using the method to move boxes in an embodiment;

[0042] Fig.12 This is a process diagram of turning the steering wheel using the method in the embodiment. DETAILED DESCRIPTION

[0043] The embodiments of the technical solution of the present invention will be described in detail below in conjunction with the accompanying drawings. Unless otherwise specified, the technical terms or scientific terms used in this application are generally understood by those skilled in the art.

[0044] Humanoid robots are becoming more and more human-like in configuration design. The structure of the trunk and lower limbs expands the operating range of the arms. At present, the focus of humanoid robot research is on the motion planning of the arms, and there is little research on the motion of the trunk of humanoid robots. In terms of mechanical configuration, there are many humanoid designs of robots, but current research is still focused on the upper and lower limbs. There is still a lack of work on separate motion planning for the trunk of humanoid upper limb robots. For humanoid robots with high redundancy, it is difficult to establish an accurate optimization objective function. Therefore, using imitation learning can avoid this modeling process and directly transfer the optimized actions of humans to robots. However, there are three difficulties in using imitation learning. The first difficulty is how to collect expert data for learning. The second difficulty is how to apply expert data to robots; the third difficulty is how to determine the characteristics of network input. The robot has a complex structure and many degrees of freedom, which leads to a large amount of feature information in the system. Input without pre-processing will greatly reduce the effect of network learning.

[0045] In view of the above, this paper studies how to generate humanoid trunk motions in the collaborative operation of humanoid upper limb robots to achieve better arm movements. Figure 1-Figure 2 ,This scheme proposes a humanoid robot trunk motion planning method,which includes the following steps:

[0046] S1. Use human skeleton recognition technology to identify human demonstration movements;

[0047] S2, construct a skeleton recognition point set;

[0048] S3, establish a mapping relationship from the Cartesian space of the skeleton recognition point set to the joint space of the robot trunk, and obtain a trunk joint angle data set of the robot; this implementation step applies the expert data to the robot through mapping, eliminating the obstacle that the collected expert data cannot be directly applied to the robot due to the configuration difference between the human body and the robot;

[0049] S4. Based on imitation learning, a neural network is used to train the relationship between the coordinate positions of the skeleton recognition points and the trunk joint angle data; this implementation step does not plan the movement of the trunk from a wide workspace and operability calculation, but directly extracts the movement mechanism from the human body motion data based on the idea of ​​imitation learning.

[0050] S5. The output obtained by the neural network training is used as the input of the second-order system model. The desired trunk joint angle and angular velocity are calculated by the second-order system model to complete the motion planning of the humanoid robot trunk. This implementation step uses the second-order system to implement the hybrid architecture of the trunk neural network and the velocity-level inverse kinematics of the arms.

[0051] Specifically, refer to Figure 3 Note that this embodiment provides a three-degree-of-freedom trunk motion planning for a humanoid robot, and the specific process is as follows:

[0052] In step S1, the human skeleton recognition technology in the ZED camera is used to set a human body demonstration action. In order to reduce the error recognition in the camera recognition process and prevent the recognized skeleton points from deviating too much from the actual situation, the median filter is used after this step to play a role in smoothing. The key points of the human skeleton in the human body movement process can be captured by recognition. For example: select the 0th, 3rd, 4th, and 11th key points, establish a mapping relationship from the skeleton points in the Cartesian space to the angle values ​​in the joint angle space, and obtain the trunk joint angle expert data set corresponding to the different robot arm movement positions of the robot trunk.

[0053] The skeleton recognition point set constructed in step S2 includes the skeleton recognition points of the human body demonstration and the skeleton recognition points obtained by randomly extracting the data set from the skeleton recognition points of the human body demonstration using the data enhancement method. Specifically: Figure 5 The figure shows the position trajectory of the left and right wrist points of the human body collected during 60 box-moving actions (the left figure is the overall figure, the middle figure is the side view, and the right figure is the top view). The blue point represents the trajectory of the left wrist point, and the green point represents the trajectory of the right wrist point. It can be seen from the figure that collecting data solely based on human demonstration has the problems of low efficiency, sparse data distribution and uneven distribution. In order to improve efficiency, the data enhancement method is used to quickly fill in the blank areas in the data collected based on human demonstration. The specific process of obtaining skeleton recognition points by the data enhancement method is as follows: randomly extract two points from the existing human demonstration skeleton point data as a group, perform linear interpolation, and obtain points divided into one quarter, one half and three quarters each time, and obtain three groups of skeleton recognition points. Repeat the random extraction times and interpolation, and finally obtain several groups of skeleton recognition points after data enhancement.

[0054] For example, suppose 10,000 sets of skeleton recognition points are obtained through human body demonstration, and then these 10,000 sets of points are randomly selected in groups of two for interpolation. Each interpolation can obtain points divided into one quarter, one half, and three quarters. That is, each interpolation operation can increase 3 sets of skeleton points. Repeat the random difference 5,000 times, and you can get 15,000 sets of skeleton recognition points (such as Figure 6The left and right wrist trajectory point diagrams obtained by data enhancement are shown in the figure. The left picture is the overall picture, the middle picture is the side view, and the right picture is the top view). These 15,000 groups are obtained by data enhancement, and the final data set contains 10,000+15,000 groups of data.

[0055] go through Figure 3 The human body demonstration, median filtering and linear interpolation in the robot are used to obtain a sufficient number of human skeleton recognition points through human body action demonstration and data enhancement. Next, the skeleton recognition point set needs to be mapped from Cartesian space to the three-degree-of-freedom joint space of the robot torso. The coordinate mapping transformation is performed using geometric calculation to establish the geometric relationship between the human skeleton recognition points 0, 3, 4, 11 and the three joints of the torso, such as Figure 7 The geometric mapping relationship diagram from Cartesian space to joint space is shown.

[0056] Specifically: Data collection and establishment of coordinate system: The ZED camera recognizes the skeleton points of the upper body, and uses skeleton points 0, 3, 4, and 11 as key points to establish the human body coordinate system as X B Y B Z B , the coordinate origin O coincides with the skeleton point 0, X B The direction from skeleton point 0 to skeleton point 22, Z B is the vertical upward direction, Y B Determined by the right-hand system, such as Figure 7 (a) Cartesian space and skeleton recognition point distribution diagram;

[0057] Coordinate mapping transformation: Angle value θ of the first torso joint t1 Defined as the line connecting skeleton points 0 and 3 and Z B The angle between directions, such as Figure 7 (b) As shown in the figure. The plane perpendicular to the line between 0 and 3 is defined as P1, and the angle value of the second torso joint is θ t2 Defined as the projection of the line connecting skeleton points 4 and 11 on P1 and X B The angle of the third trunk joint is θ t3 It is defined as the angle between the line connecting skeleton points 4 and 11 and the projection of this line on P1, such as Figure 7 (c) as shown.

[0058] Through the mapping transformation in step S3, the joint angle expert dataset of the three-degree-of-freedom robot trunk is obtained. In the following step S4, based on the idea of ​​action cloning in imitation learning, a fully connected neural network is used to learn the relationship between the coordinate positions of skeleton points 7 and 14 and the trunk joint angle data in step S3. The trunk joint angle values ​​required for different wrist point positions can be obtained through the neural network, thereby achieving the motion planning effect of the trunk adaptive double-arm position.

[0059] In step S5, the output of the neural network learning is used as the input of the second-order system model to obtain the desired trunk joint angle vector θ t and the angular velocity vector where Θ t =(θ t1 ,θ t2 ,θ t3 ), The second-order system model is shown below:

[0060]

[0061] In the formula, the subscript i represents the i-th joint of the trunk, M i represents the system quality of the second-order system, K di represents the damping coefficient of the second-order system, K si represents the stiffness coefficient of the second-order system, θ ti , and represent the desired trunk joint angle, angular velocity and angular acceleration respectively, and F ei The system external force input of the second-order system is equal to the network output value learned by the neural network. The above formula can be used to obtain the expected joint angle and angular velocity after the second-order system is processed.

[0062] The feasibility and advantages of the trunk motion planning method of this scheme are verified by an embodiment below.

[0063] Raising the arm is a common action for humanoid robots. When the robot is in the initial working state, in order to maintain the balance of the body and reduce the load on the trunk joints, the trunk is often kept vertical and the arms are in a drooping state. After entering the working state, the robot needs to perform an operation task, and in this process, the arm needs to be raised to bring the end effector close to the object being operated.

[0064] The torso motion planning method (TMPNet) proposed in this application is compared with the damped least squares method (DLSM) and the fixed torso method (FixedMethod) (the three methods are simplified in Figures 9(a), 9(b), 9(c), 10(a), 10(b) and 10(c)). The advantages of the proposed method are reflected by comparing the movement of the torso during the arm raising action.

[0065] The planned arm-lifting action is for the left and right arms to reach the target points (-0.25, 0.9, 0.5) and (0.25, 0.9, 0.5) from the initial positions respectively. Figure 8 (a) Figure 8 (b) and Figure 8 (c) shows the overall motion of the robot in three cases: the trunk motion planning method of this scheme, the damped least squares method, and the fixed trunk. The blue line is the motion trajectory of the ends of the two arms. Figures 9(a), 9(b), and 9(c) are curves showing the angles and angular velocities of the three joints of the trunk changing with time when the above three methods are used in turn. It can be seen that when the damped least squares method is used, the first joint of the trunk will rotate backward and then forward, while when the trunk motion planning network is used, the trunk will directly move forward to approach the target. Since the damped least squares method does not distinguish between the trunk joints and the arm joints, compared with the trunk motion planning method of this scheme, the movement amplitude of the trunk joint 1 of the former is also larger, indicating that the energy consumption is relatively high.

[0066] In comparison with the maneuverability of the arms, the flexibility of the robot arm, that is, the ability to move easily in any direction, can be called its maneuverability. It is a scalar measure, higher is better, and can be calculated for each point on the trajectory through the Jacobian matrix. Figures 10(a) and 10(b) are graphs showing the change in maneuverability of the arms over time in each of the three methods. It can be seen that when the trunk motion planning method of this scheme and the traditional damped least squares method are used, the change in the trunk joint angle can keep the ends of the arms at a high maneuverability, which is convenient for the arms to perform task operations, while the fixed trunk will result in lower maneuverability, which is not conducive to the flexible operation of the robot.

[0067] Finally, the energy consumption of the three methods is compared. Figure 10(c) shows the energy consumption curve during the arm lifting process over time. The curve is divided into three processes. First, the time period from 0 to 5 seconds indicates that the robot is in the initial state, 5 to 15 seconds is the arm lifting process, and 15 to 30 seconds is the stabilization process after reaching the target point. It can be seen from the figure that during the arm lifting time of 5 to 15 seconds, the total current value is always at a low level when the method proposed in this scheme is used. Therefore, it can be obtained that the method proposed in this scheme can minimize the overall energy consumption of the robot. The less energy consumption, the better, and the better energy saving.

[0068] The following is an application verification of this method:

[0069] Combining the above trunk motion planning method with the decoupled damped least squares method, the decoupled damped least squares method can achieve the given trunk joint angular velocity command. In this case, the target tracking of the dual arms can realize the task operation of the humanoid upper limb robot. The formula of the decoupled damped least squares method is as follows:

[0070]

[0071] In the formula, the subscript k represents the left arm or the right arm, represents the joint angular velocity vector of the left and right arms, is the pseudo-inverse matrix of the Jacobian matrix of the left and right arms, is the velocity vector of the left and right arms tracking the target, X r To track the position and attitude vector of the target, X k represents the position and attitude vector of the left and right end-effectors, J k，t represents the Jacobian matrix between the forward kinematics of the dual-arm end effector and the torso joint, is the trunk motion command obtained by the above trunk motion planning method, K p is the proportionality coefficient.

[0072] The method proposed in this application is applied to a humanoid upper limb robot entity to perform task operations. The robot can complete tasks such as box carrying and steering wheel turning. The task operation process is as follows Fig.11 and Fig.12 shown.

[0073] The present invention has been disclosed as above with preferred implementation cases, but it is not used to limit the present invention. Any technician familiar with the profession can make slight changes or modifications to equivalent implementation cases with equivalent changes by using the above-disclosed structures and technical contents without departing from the scope of the technical solution of the present invention, which still fall within the scope of the technical solution of the present invention.

Claims

1. A method for planning the motion of a humanoid robot trunk, characterized in that: The following steps are involved: S1. Use human skeleton recognition technology to identify human demonstration movements; S2, construct a skeleton recognition point set; S3, establishing a mapping relationship from the Cartesian space of the skeleton recognition point set to the joint space of the robot trunk, and obtaining a trunk joint angle data set of the robot; The process of establishing the mapping relationship from the Cartesian space of the skeleton recognition point set to the joint space of the robot trunk is as follows: Data collection and coordinate system establishment: The ZED camera identifies the skeleton points of the upper body, and uses skeleton points 0, 3, 4, and 11 as key points to establish the human body coordinate system as X B Y B Z B , the coordinate origin O coincides with the skeleton point 0, X B The direction from skeleton point 0 to skeleton point 22, Z B is the vertical upward direction, Y B Determined by the right hand system; Coordinate mapping transformation: Angle value θ of the first torso joint t1 Defined as the line connecting skeleton points 0 and 3 and Z B The angle between the directions, the plane perpendicular to the line between 0 and 3 is defined as P1, the angle value of the second torso joint is θ t2 Defined as the projection of the line connecting skeleton points 4 and 11 on P1 and X B The angle of the third trunk joint is θ t3 Defined as the angle between the line connecting skeleton points 4 and 11 and the projection of this line on P1; S4, using a neural network to train the relationship between the coordinate positions of the skeleton recognition points and the trunk joint angle data based on imitation learning; S5. The output obtained from the neural network training is used as the input of the second-order system model. The desired trunk joint angle and angular velocity are calculated by the second-order system model to complete the motion planning of the humanoid robot trunk.

2. A method for planning humanoid robot trunk motion according to claim 1, characterized in that: In step S1, a ZED camera is used to recognize human demonstration movements.

3. The method for humanoid robot trunk motion planning according to claim 1, characterized in that: The skeleton recognition point set of step S2 includes skeleton recognition points of human body demonstration and skeleton recognition points obtained by randomly extracting a data set from the skeleton recognition points of human body demonstration using a data enhancement method.

4. A humanoid robot trunk motion planning method according to claim 3, characterized in that: The specific process of obtaining skeleton identification points by the data enhancement method is as follows: randomly extract two points as a group from the existing human body demonstration skeleton point data, perform linear interpolation, and obtain points divided into one quarter, one half, and three quarters each time to obtain three groups of skeleton identification points, repeat the random extraction times and interpolation, and finally obtain several groups of skeleton identification points after data enhancement.

5. The method for planning the motion of a humanoid robot trunk according to claim 1, characterized in that: In step S4, a fully connected neural network is used to learn the relationship between the coordinate positions of skeleton points 7 and 14 and the trunk joint angle data in step S3.

6. The method for humanoid robot trunk motion planning according to claim 1, characterized in that: The specific process of obtaining the desired trunk joint angle and angular velocity from the output of the second-order system model in step S5 is: The second-order system model is expressed as follows: In the formula, the subscript i represents the i-th joint of the trunk, M i represents the system quality of the second-order system, K di represents the damping coefficient of the second-order system, K si represents the stiffness coefficient of the second-order system, θ ti , and represent the desired trunk joint angle, angular velocity and angular acceleration respectively, and F ei Represents the system external force input of the second-order system, which is equal to the network output value learned by the neural network.

Images