Method, application and equipment for generating anthropomorphic foot end track of biped robot based on full-connection network

Abstract

The invention belongs to the related technical field of biped robot gait planning, and discloses a method, application and equipment for generating a biped robot anthropomorphic foot end trajectory based on a full-connection network, and the method comprises the steps: (1) grouping the collected foot end trajectory data according to a starting step, an intermediate circulation step and a termination step; (2) mapping each group of foot end track data into foot end track data which can be used by a robot based on a kinematics mapping relation between a human body and the robot, and carrying out conformal interpolation; (3) training a full-connection neural network model by adopting the interpolated foot end trajectory data, further generating a foot end trajectory of each intermediate step by adopting the full-connection neural network model, and mapping the directly generated foot end trajectories of the starting step and the ending step to jointly form an anthropomorphic foot end trajectory; and the main loss of back propagation of the full-connection neural network model is punished by introducing the trajectory smoothness and the prediction precision of the start end and the tail end. According to the method, the anthropomorphism of the foot end track is improved.

Description

Technical Field

[0001] This invention belongs to the technical field of bipedal robot gait planning, and more specifically, relates to a method, application, and device for generating anthropomorphic foot trajectories of bipedal robots based on fully connected networks. Background Technology

[0002] Bipedal robots interact with their environment discretely through their legs during movement, resulting in significant nonlinearity in their system dynamics equations, posing a major challenge to gait control. Based on the natural oscillation of the center of mass during human walking, this paper focuses on generating the robot's center of mass trajectory and uses a simplified model to describe the dynamic state of the bipedal robot's center of mass. This effectively reduces the coupling of the robot's dynamic parameters and is currently widely used in the field of bipedal robot gait planning.

[0003] The gait planning method based on a simplified model treats the bipedal robot as an inverted pendulum model, such as a linear inverted pendulum model or a spring-loaded inverted pendulum model. It assumes that the total mass of the robot is concentrated at the center of mass. The dynamic characteristics of the center of mass of the bipedal robot are described by the dynamic equations of the simplified model. Then, Model Predictive Control (MPC) is used to plan the trajectory of the robot's center of mass. Finally, the trajectory of the foot end is generated by Bézier curves, thereby realizing the gait planning of the bipedal robot.

[0004] The COT (Cost of Transport) parameter for bipedal robots is defined as the ratio of energy consumption to distance traveled and body weight, and is one of the core indicators for measuring the anthropomorphism of a robot. Gait planning for humanoid robots mainly includes planning the robot's center of mass trajectory and foot trajectory planning. Traditional model-based gait planning methods focus on planning the robot's center of mass trajectory, while foot trajectory planning typically involves simply using Bézier curves. This results in traditional gait planning methods where the foot trajectory lacks anthropomorphic characteristics, failing to follow the biomimetic characteristics of toe-off and heel-on-the-ground movements, leading to stiff gait behavior and high energy consumption. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method, application, and device for generating anthropomorphic foot trajectories for bipedal robots based on fully connected networks. This aims to solve the problem that existing gait planning methods lack anthropomorphic characteristics in foot trajectories.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for generating anthropomorphic foot trajectories of a bipedal robot based on a fully connected network is provided, the method comprising the following steps:

[0007] (1) The foot trajectory data collected by the human motion capture device during normal walking is grouped according to the starting step, intermediate cycle step and ending step.

[0008] (2) Based on the kinematic mapping relationship between the human body and the robot, each set of foot trajectory data is mapped to foot trajectory data that the robot can use, and the mapped foot trajectory is subjected to conformal interpolation to a fixed length;

[0009] (3) The interpolated foot trajectory data is used to train the fully connected neural network model, and then the trained fully connected neural network model is used to generate the foot trajectory of each intermediate step. The foot trajectory of each intermediate step, the foot trajectory of the starting step and the foot trajectory of the ending step directly generated based on the foot trajectory mapping together form the anthropomorphic foot trajectory. During training, the trajectory smoothness and the prediction accuracy of the start and end points are introduced to penalize the main loss of the backpropagation of the fully connected neural network model.

[0010] Furthermore, the input dimension of the fully connected neural network model is Output dimension is The system has four fully connected hidden layers with 16, 64, 256, and 64 neurons in each layer, respectively, and uses ReLU as the activation function.

[0011] Furthermore, the local prediction accuracy loss is defined as:

[0012]

[0013] Focus on the prediction accuracy of the initial k points and the last m points of the trajectory, Y t Let t be the t-th point on the predicted foot trajectory; Let t be the t-th point on the actual foot trajectory; L is the sequence length in each dimension.

[0014] Furthermore, the trajectory smoothness loss is defined as:

[0015]

[0016] Furthermore, the main loss adopts mean squared error (MES), and the total loss obtained by weighted summation of the main loss, local prediction accuracy loss and trajectory smoothing loss is used for backpropagation.

[0017] Furthermore, the smoothness of the trajectory is penalized by the first and second differences of the trajectory.

[0018] Furthermore, the length of each set of foot trajectory is shape-preservingly interpolated to length L. max L max This represents the maximum sampling length for all intermediate loop steps.

[0019] The present invention also provides a gait planning method for a bipedal robot, the gait planning method comprising the following steps:

[0020] S1. The anthropomorphic foot trajectory of the bipedal robot is obtained by using the method described above based on a fully connected network to generate the anthropomorphic foot trajectory of the bipedal robot. The obtained anthropomorphic foot trajectory is resampled so that the foot trajectory sampling frequency is consistent with the control frequency of the robot. The resampled foot trajectory is parameterized and post-processed using a high-order Bezier curve so that the start and end points of the foot trajectory coincide with the front and rear foot landing points, and the start and end point velocities are 0.

[0021] S2, based on linear MPC, performs real-time rolling planning of the center-of-mass trajectory of the bipedal robot in the sagittal and coronal planes, thereby realizing gait planning of the bipedal robot.

[0022] The present invention also provides an electronic device, the electronic device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program executing the method described above for generating anthropomorphic foot trajectory of a bipedal robot based on a fully connected network or the gait planning method of a bipedal robot described above.

[0023] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the method for generating anthropomorphic foot trajectories of a bipedal robot based on a fully connected network as described above, or the gait planning method for a bipedal robot as described above.

[0024] In summary, compared with the prior art, the method, application, and device for generating anthropomorphic foot trajectories of bipedal robots based on fully connected networks provided by the present invention have the following beneficial effects:

[0025] 1. The fully connected neural network model is trained using interpolated foot trajectory data, and then the trained fully connected neural network model is used to generate the foot trajectory for each intermediate step. The method of generating robot foot trajectory using fully connected neural network can not only effectively realize the anthropomorphic foot movement mode of heel-first and toe-lagging, but also significantly improve the bio-kinematic similarity characteristics of the swing phase trajectory curve in three-dimensional space.

[0026] 2. By applying trajectory smoothness penalty and initial and final local prediction accuracy penalty to the loss of the fully connected neural network, and by performing parameterized post-processing on the generated trajectory, the impact of the robot's foot on the ground is reduced, and the stability of the foot touching the ground is improved.

[0027] 3. Based on a fully connected neural network method, real-time foot trajectory generation with biomimetic characteristics is realized in robot gait planning, which can achieve a lower COT value and make the robot's lower limb movement more energy efficient.

[0028] 4. Perform conformal interpolation on the length of each foot trajectory to the length L. max (L max (The maximum sampling length for all intermediate loop steps) is used to adapt to fully connected neural networks.

[0029] 5. Smoothness of the trajectory: By penalizing the first and second order differences of the trajectory, the swinging foot trajectory generated by the model is made smoother, reducing foot shaking.

[0030] 6. The resampled foot trajectory is parameterized and post-processed using a high-order Bézier curve to ensure that the starting and ending points of the foot trajectory coincide with the positions of the front and rear landing points to ensure kinematic continuity, and the starting and ending point velocities are 0 to ensure no impact between the foot and the ground. Attached Figure Description

[0031] Figure 1 This is a flowchart of gait planning for a bipedal robot provided in an embodiment of the present invention;

[0032] Figure 2 This is a diagram showing the intermediate loop step size and speed distribution provided in an embodiment of the present invention;

[0033] Figure 3 This is a mapping diagram of the human body and robot foot trajectory involved in the embodiments of the present invention;

[0034] Figure 4 This is a schematic diagram of the loss penalty mechanism involved in an embodiment of the present invention;

[0035] Figure 5 This is a convergence curve of the FCNN training loss provided in an embodiment of the present invention;

[0036] Figure 6 (a), (b), and (c) in the figure are the foot spatial position trajectory curve, the foot yaw posture curve, and the foot pitch posture curve, respectively.

[0037] Figure 7 These are gait simulation frames from an embodiment of the present invention;

[0038] Figure 8 This is a COT curve diagram of an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0040] This invention provides a method for generating anthropomorphic foot trajectories for bipedal robots based on fully connected neural networks. It learns and generates anthropomorphic foot trajectories for bipedal robots in real time based on fully connected neural networks (FCNN), replacing the simple planning of robot foot trajectories using Bézier curves. This achieves more human-like gait planning and has a lower COT value compared to traditional methods.

[0041] Please see Figure 1 The method mainly includes the following steps:

[0042] Step 1: Use human motion capture equipment to collect foot trajectory data during normal walking, and group the foot trajectory data according to the starting step, intermediate cycle step and ending step.

[0043] To analyze the kinematic characteristics of different stages of normal human walking, the stepping phase was divided into three stages: the starting step, the intermediate cycle step, and the ending step. During sampling, the human body performed normal forward walking in a straight line with different turning radii and small stride forward walking at different speeds on flat terrain.

[0044] Based on the Xsens motion capture suit, with a sampling frequency of 240Hz, a total of 1,575,230 valid data points were collected. The data were grouped according to each step, and each group contained the sequence of the swinging foot in five dimensions of three-dimensional space: X, Y, Z, Yaw, and Pitch (the origin of the coordinate system is the departure point of the swinging foot; during the walking process, the human foot does not have the Roll degree of freedom, so this angle is not considered and is always regarded as 0).

[0045] The collected data were grouped. The sequence length L for each dimension was the duration of each step multiplied by the sampling frequency, resulting in a total of 1517 data groups, including 33 groups each for the start and end steps, and 1451 groups for intermediate loop steps. The input for each data group was the relative position [Δx, Δy, Δz, Δyaw, Δpitch] of the swinging foot at the landing and departure points (obtained by subtracting the beginning of the sequence from the end of the sequence), and the output was the trajectory sequence P = [p X ,p Y ,p Z ,p Yaw ,p PitchThe swing trajectory of a human swing foot is usually determined by the relative position of the landing point and the departure point. That is, the relative pose describes the overall characteristics of the swing trajectory. Therefore, theoretically, a 5*L-dimensional trajectory sequence can be generated by predicting the 5-dimensional relative pose.

[0046] In the dataset, the step length of the intermediate loop steps ranges from 0.03m to 1.4m, and the velocity ranges from 0.015m / s to 1.75m / s. Furthermore, the step length and velocity exhibit a strong linear correlation. Figure 2 As shown.

[0047] The step length and velocity distribution of the starting and ending steps are relatively concentrated. The step length of the starting step is concentrated around 0.35m, and the velocity distribution is concentrated around 0.3m / s; the step length of the ending step is concentrated around 0.5m, and the velocity distribution is concentrated around 0.42m / s.

[0048] Step 2: Based on the kinematic mapping relationship between the human body and the robot, each set of foot trajectory data is mapped to foot trajectory data that the robot can use, and the mapped foot trajectory is subjected to conformal interpolation to a fixed length so that the discrete point data in each foot trajectory segment are equal.

[0049] Specifically, because the length of the lower limbs of a bipedal robot is different from that of a human, the foot trajectories captured by motion capture equipment cannot be directly applied to bipedal robots. Therefore, it is necessary to map the captured foot trajectories based on the limb length ratio to adapt them to the robot.

[0050] Assuming that at least one of the operator's feet is in contact with the ground during the data collection of human foot trajectory, and that the foot does not slip during walking, the swinging foot trajectory can be scaled according to the ratio of the robot's leg length to the human's leg length:

[0051]

[0052] Where K = diag(k) x ,k y ,k z (1,1) is a diagonal matrix representing the scaling relationship, k x k y and k z Let be the scaling factor, here we set The operator's height was measured to be 1.67m, and the lower leg length was... Thigh length The experimental platform uses Tiangong v1.1pro, and the robot's lower leg and upper leg lengths are [not specified]. The mapping relationship between human body and robot foot trajectory is as follows: Figure 3 As shown.

[0053] Spatial position of foot trajectory [p] X ,p Y ,p Z Mapping is required, and the attitude angle [p] Yaw ,p Pitch The remainder will remain unchanged.

[0054] Because the duration of each step varies during actual data collection, each set of foot trajectory sequences P... i The lengths are not the same. Fully connected neural networks are not good at handling variable-length output data, so the length of each foot trajectory is shape-preservingly interpolated to a length L. max (L max (The maximum sampling length for all intermediate loop steps).

[0055] Step 3: The interpolated foot trajectory data is used to train the fully connected neural network model. Then, the trained fully connected neural network model is used to generate the foot trajectory for each intermediate step. The foot trajectory for each intermediate step, the foot trajectory for the starting step and the foot trajectory for the ending step directly generated based on the human foot trajectory mapping, together form the anthropomorphic foot trajectory. During training, trajectory smoothness (quantized using the first and second differences of the foot trajectory) and the prediction accuracy of the start and end points are introduced to penalize the main loss (mean square error of the overall trajectory) of the backpropagation of the fully connected neural network model.

[0056] Specifically, for the intermediate loop step, assuming the robot's target foot landing point is known, the fully connected neural network model needs to realize the swing foot trajectory sequence P = [p] based on the relative poses [Δx, Δy, Δz, Δyaw, Δpitch] of two adjacent target foot landing points. X ,p Y ,p Z ,p Yaw ,p Pitch [Prediction].

[0057] Construct a fully connected neural network model with input dimension . Output dimension is The system consists of four fully connected hidden layers with 16, 64, 256, and 64 neurons per layer, respectively. The activation function is ReLU, and regularization is performed using Dropout = 0.01.

[0058] In the training of the FCNN model, the backpropagation loss is penalized by the trajectory smoothness and the local prediction accuracy of the start and end positions, such as... Figure 4 As shown.

[0059] The local prediction accuracy loss is defined as:

[0060]

[0061] We focus on the prediction accuracy of the initial k points and the last m points of the trajectory (in this implementation, we take k = m = 20). This penalty can make the swing foot trajectory generated by the model closer to the predetermined target point at the foot lift and foot landing points.

[0062] The trajectory smoothness loss is defined as:

[0063]

[0064] The smoothness of the trajectory is improved by penalizing the first and second order differences of the trajectory, making the swinging foot trajectory generated by the model smoother and reducing foot shaking.

[0065] The main loss is calculated using Mean Squared Error (MES), and the final backpropagation uses the weighted sum of the three losses. After 290 training iterations, the model achieves a good result. After 295 training iterations, the model achieves a good result, as shown in the training curve. Figure 5 As shown.

[0066] The present invention also provides a gait planning method for a bipedal robot, the gait planning method comprising the following steps:

[0067] S1. The anthropomorphic foot trajectory of the bipedal robot is obtained by using the method described above based on a fully connected network to generate the anthropomorphic foot trajectory of the bipedal robot. The obtained anthropomorphic foot trajectory is resampled to make the foot trajectory sampling frequency consistent with the robot's control frequency. The resampled foot trajectory is parameterized and post-processed using a high-order Bézier curve to make the start and end points of the foot trajectory coincide with the front and rear foot landing points to ensure kinematic continuity. The start and end point velocities are 0 to ensure no impact between the foot and the ground.

[0068] Specifically, since the sampling frequency of the bipedal robot's swinging foot trajectory needs to be consistent with the robot's control frequency, if the robot is controlled at a frequency of 100Hz, the foot trajectory sequence generated by FCNN needs to be resampled to 100Hz, that is, the trajectory points are interpolated or cropped into a new sequence with equal time intervals δ = 0.01s on the original time series.

[0069] To ensure robot gait stability, the planned robot foot trajectory must match the pose of the target foot landing point at both the departure and landing points, and the velocity must be zero at the moment of departure and touchdown. However, the foot trajectory generated by FCNN cannot directly satisfy these two hard constraints. Therefore, this implementation method parameterizes the foot trajectory generated by FCNN. The specific method is as follows:

[0070] (1) The generated foot trajectory is approximated using a higher-order Bézier curve, where t is the normalized time coefficient, n is the order of the Bézier curve (n = 8 in this embodiment, but can be varied depending on the situation), and A is the n+1 curve control points; the expression for the higher-order Bézier curve is:

[0071]

[0072] (2) If the pose of the previous and current target landing points is and Then constraint B(0) = 0,

[0073] (3) Constraints Solve for the remaining n-3 control points.

[0074] After resampling and trajectory parameterization, the foot trajectory can be used for robot gait control.

[0075] Regarding the starting and ending steps, considering that the human foot trajectory dataset covers various scenarios of normal forward walking, regardless of speed or turning radius, the foot trajectory for each starting step and ending step is quite similar. Similarly, the robot's starting and ending steps can be predetermined in advance. This involves directly selecting representative starting and ending step trajectories from the dataset, resampling and parameterizing them, and then using them directly as the swinging foot trajectory for each starting and ending step, without needing to generate them through a fully connected neural network. Figure 6 As shown.

[0076] S2, based on linear MPC, performs real-time rolling planning of the center-of-mass trajectory of the bipedal robot in the sagittal and coronal planes, thereby realizing gait planning of the bipedal robot.

[0077] Specifically, after completing the foot trajectory planning, in order to achieve complete gait planning, the next step is to perform real-time rolling planning of the robot's center of mass trajectory in the sagittal and coronal planes based on linear MPC. The following example uses the center of mass trajectory in the sagittal plane:

[0078] The minimum time interval is δ. Based on the linear inverted pendulum model to describe the robot's center of mass state, the robot's center of mass state and zero torque point position at the k-th time are denoted as δ. z k =x ZMP (kδ), the discretized state transition equation is:

[0079]

[0080] z k=[1 0-h / g]x k

[0081] In the formula, h and g are the robot's center of mass height and gravitational acceleration, respectively. The prediction window length for model predictive control rolling optimization is N (i.e., predictive control is performed over the next Nδ time intervals), denoted as... Based on the recursive relationship of the state transition equation, a quadratic optimization problem is established, with the cost function and constraints as follows:

[0082]

[0083]

[0084] Where Q and R are constants used to balance the ZMP tracking error with the oscillations of the centroid acceleration derivative. To support the lower bound of the domain, This is the upper limit of the support domain (equal to the anterior and posterior boundaries of the support foot in the sagittal plane). When the sequence of target landing point positions is known, the reference position of the ZMP is... Given, it can be obtained Substituting the first element of the sequence into the discretized state transition equation yields the target centroid state at the next moment.

[0085] The present invention also provides an electronic device, the electronic device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program executing the method described above for generating anthropomorphic foot trajectory of a bipedal robot based on a fully connected network or the gait planning method of a bipedal robot described above.

[0086] The present invention will be further described in detail below with reference to specific embodiments.

[0087] Using the fully connected neural network model and gait planning method of this invention, gait planning was performed on the Tiangong v1.1pro humanoid robot. The robot's total mass is 51.6 kg, the length of its thigh and lower leg are both 0.3 m, the center of mass height when standing still on a flat terrain is set to 0.72 m, and the length and width of its feet are 0.18 m and 0.1 m, respectively. Simulation experiments were conducted in Webots with a control frequency of 100 Hz, i.e., a minimum time interval δ = 0.01 s.

[0088] Figure 7 The diagram shows gait simulation frames implemented by applying the gait planning method proposed in this invention to the Tiangong v1.1pro robot. This gait effectively realizes an anthropomorphic foot movement pattern with heel-first contact and toe-removal lag, while significantly improving the bio-kinematic similarity characteristics of the swing foot trajectory curve in three-dimensional space.

[0089] Figure 8 The COT curve for the robot's walking process is shown below. In this example, the COT is calculated using the following formula (since energy losses such as joint friction and heat generation can be ignored in the simulation environment):

[0090]

[0091] In the formula τ i and Let represent the torque and angular velocity of the i-th joint, Mg be the robot's total gravity, and v be the robot's center of mass velocity. The changes in COT (Cost-to-Occupation Time) for two walking modes were compared: one using FCNN-generated foot trajectories, and the other using Bézier curves. The average COT for the walking mode using Bézier curves to generate foot trajectories was 0.4999, while the average COT for the walking mode using FCNN-generated foot trajectories was 0.4544, a decrease of 9.1%. This reflects that the gait planning method based on FCNN-generated foot trajectories proposed in this invention has better anthropomorphism and superior energy efficiency.

[0092] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the method for generating anthropomorphic foot trajectories of a bipedal robot based on a fully connected network as described above, or the gait planning method for a bipedal robot as described above.

[0093] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating anthropomorphic foot trajectories of a bipedal robot based on a fully connected network, characterized in that, The method includes the following steps: (1) The collected foot trajectory data during normal human walking are grouped according to the starting step, intermediate cycle step and ending step; (2) Based on the kinematic mapping relationship between the human body and the robot, each set of foot trajectory data is mapped to foot trajectory data that the robot can use, and the mapped foot trajectory is subjected to conformal interpolation to a fixed length; (3) The interpolated foot trajectory data is used to train the fully connected neural network model, and then the trained fully connected neural network model is used to generate the foot trajectory of each intermediate step. The foot trajectory of each intermediate step, the foot trajectory of the starting step and the foot trajectory of the ending step directly generated based on the foot trajectory mapping together form the anthropomorphic foot trajectory. During training, the trajectory smoothness and the prediction accuracy of the start and end points are introduced to penalize the main loss of the backpropagation of the fully connected neural network model.

2. The method for generating anthropomorphic foot trajectories of a bipedal robot based on a fully connected network as described in claim 1, characterized in that: The input dimension of a fully connected neural network model is Output dimension is The system has four fully connected hidden layers with 16, 64, 256, and 64 neurons in each layer, respectively, and uses ReLU as the activation function.

3. The method for generating anthropomorphic foot trajectories of a bipedal robot based on a fully connected network as described in claim 1, characterized in that: The local prediction accuracy loss is defined as: Focus on the prediction accuracy of the initial k points and the last m points of the trajectory, Y t Let t be the t-th point on the predicted foot trajectory; Let t be the t-th point on the actual foot trajectory; L is the sequence length in each dimension.

4. The method for generating anthropomorphic foot trajectories of a bipedal robot based on a fully connected network as described in claim 3, characterized in that: The trajectory smoothness loss is defined as:

5. The method for generating anthropomorphic foot trajectories of a bipedal robot based on a fully connected network as described in claim 4, characterized in that: The main loss is calculated using mean squared error (MES), and the total loss obtained by weighted summation of the main loss, local prediction accuracy loss, and trajectory smoothing loss is used for backpropagation.

6. The method for generating anthropomorphic foot trajectories of a bipedal robot based on a fully connected network as described in any one of claims 1-5, characterized in that: The smoothness of the trajectory is penalized by the first and second differences of the trajectory.

7. The method for generating anthropomorphic foot trajectories of a bipedal robot based on a fully connected network as described in any one of claims 1-5, characterized in that: The length of each set of foot trajectory is shape-preservingly interpolated to length L. max L max This represents the maximum sampling length for all intermediate loop steps.

8. A gait planning method for a bipedal robot, characterized in that, The gait planning method includes the following steps: S1, the method for generating bipedal robot anthropomorphic foot trajectory based on fully connected network as described in any one of claims 1-7 is used to obtain the anthropomorphic foot trajectory of the bipedal robot. The obtained anthropomorphic foot trajectory is resampled so that the foot trajectory sampling frequency is consistent with the robot's control frequency. The resampled foot trajectory is parameterized and post-processed using a high-order Bezier curve so that the start and end points of the foot trajectory coincide with the front and rear foot landing points, and the start and end point velocities are 0. S2, based on linear MPC, performs real-time rolling planning of the center-of-mass trajectory of the bipedal robot in the sagittal and coronal planes, thereby realizing gait planning of the bipedal robot.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it executes the method for generating anthropomorphic foot trajectory of a bipedal robot based on a fully connected network as described in any one of claims 1-7, or the gait planning method for a bipedal robot as described in claim 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the method for generating anthropomorphic foot trajectories of a bipedal robot based on a fully connected network as described in any one of claims 1-7, or the gait planning method for a bipedal robot as described in claim 8.

Images