Gait generation method for body-equipped intelligent robot and electronic device

Abstract

The present application relates to a kind of body intelligent robot gait generation method and electronic equipment, comprising: constructing gait rhythm generation model;Obtain the hip joint angle data of subject at target step speed, adjust the rhythm oscillation angular frequency of gait rhythm generation model according to the hip joint angle data at target step speed, obtain the gait rhythm generation model of adaptation target step speed;According to the gait rhythm generation model of adaptation target step speed, generate gait phase feature vector at target step speed;Construct the nonlinear mapping model of gait phase feature to knee joint motion feature, input gait phase feature vector at target step speed into nonlinear mapping model, output knee joint angle trajectory at target step speed.It belongs to the technical field of intelligent robot, solves the problem that gait rhythm generation model cannot realize the gait generation of cross step speed in prior art, effectively improves the speed self-adaptive ability, cross step speed generalization performance and engineering practicability of body intelligent robot gait.

Description

Technical Field

[0001] This invention relates to the field of intelligent robot technology, and in particular to a method for generating the gait of an embodied intelligent robot and an electronic device thereof. Background Technology

[0002] The Central Pattern Generator (CPG) is a self-organizing biological neural circuit widely found in the central nervous systems of vertebrates and invertebrates. It can autonomously generate stable rhythmic neural impulses without requiring peripheral rhythmic sensory feedback or continuous commands from higher brain centers, making it a core functional unit for the regulation of biological rhythmic movement. CPGs possess key advantages such as rhythmic robustness, environmental adaptability, and low computational energy consumption. They form the neural basis for instinctive rhythmic behaviors such as walking, swimming, and breathing, and have become a core theoretical model in fields such as gait generation for legged bionic robots, motor control of rehabilitation exoskeletons, and motor function reconstruction after spinal cord injury.

[0003] The core of CPG is driven by nonlinear differential equations, with the circular Hopf oscillator serving as a classic CPG implementation model. Due to its good stability, smoothness, and parameter adjustability, it is widely used for gait rhythm generation. This model can generate stable circular limiting cycles, characterize gait phase through two-dimensional state variables, and describe the intrinsic rhythmic characteristics of periodic joint movements during gait generation.

[0004] However, existing CPG gait generation technologies based on Hopf oscillators have the following technical drawbacks: most existing CPG rhythm generation models are designed for scenarios with fixed walking speeds. When the walking speed changes, data needs to be collected again for the new walking speed and the model needs to be retrained. This makes it impossible to achieve gait generalization generation across walking speeds, thus limiting the generalization performance and engineering practicality of the model. Summary of the Invention

[0005] Based on the above analysis, the embodiments of the present invention aim to provide a gait generation method and electronic device for an embodied intelligent robot, so as to solve the problem that the gait rhythm generation model in the prior art is designed for a fixed walking speed and cannot realize gait generation across walking speeds.

[0006] On one hand, embodiments of the present invention provide a gait generation method for an embodied intelligent robot, comprising: constructing a gait rhythm generation model; wherein the gait rhythm generation model is a biomimetic continuous periodic rhythm generation central model, used to generate gait phase feature vectors; acquiring hip joint angle data of a subject at a target walking speed, adjusting the rhythmic oscillation angular frequency of the gait rhythm generation model according to the hip joint angle data at the target walking speed to obtain a gait rhythm generation model adapted to the target walking speed; generating gait phase feature vectors at the target walking speed according to the gait rhythm generation model adapted to the target walking speed; constructing a nonlinear mapping model from gait phase features to knee joint motion features, inputting the gait phase feature vectors at the target walking speed into the nonlinear mapping model, and outputting the knee joint angle trajectory at the target walking speed, so as to realize gait generation of the embodied intelligent robot.

[0007] Furthermore, a gait rhythm generation model is constructed, including: acquiring hip joint angle data of the subject at a reference walking speed; calculating the rhythmic oscillation angular frequency of the gait rhythm generation model based on the hip joint angle data at the reference walking speed; determining other rhythmic parameters of the gait rhythm generation model besides the rhythmic oscillation angular frequency, with the optimization objective of minimizing the deviation between the gait phase feature vector generated by the gait rhythm generation model and the hip joint angle data at the reference walking speed; and constructing a biomimetic continuous periodic rhythm generation central model based on the other rhythmic parameters and the rhythmic oscillation angular frequency, as the gait rhythm generation model.

[0008] Furthermore, the rhythmic oscillation angular frequency is calculated as follows: a sliding window Fourier transform is performed on the hip joint angle data to obtain the spectral information corresponding to each data window; the instantaneous oscillation frequency of each data window is calculated based on the spectral information; the instantaneous oscillation frequency of each data window is smoothed to obtain the smoothed oscillation frequency, and the smoothed oscillation frequency is converted into the rhythmic oscillation angular frequency.

[0009] Furthermore, the instantaneous oscillation frequency of each data window is calculated based on the spectral information, including: determining the fundamental frequency peak index of the corresponding data window according to the spectral peak intensity and harmonic relationship of each data window; and calculating the instantaneous oscillation frequency of each data window based on the fundamental frequency peak index and the sampling parameters of the sliding window Fourier transform.

[0010] Furthermore, the instantaneous oscillation frequency of each data window is calculated based on the fundamental frequency peak index and the sampling parameters of the sliding window Fourier transform, including: the instantaneous oscillation frequency of each data window is calculated using the following formula: Where m is the sequence number of the data window, The instantaneous oscillation frequency, Index for fundamental frequency peak. is the sampling frequency of the sliding window Fourier transform, and N is the number of periodic samples in the sliding window Fourier transform.

[0011] Furthermore, a gait phase feature vector at the target gait speed is generated based on a gait rhythm generation model adapted to the target gait speed, including: acquiring knee joint angle data of the subject at the target gait speed; constructing an initial gait phase based on hip joint angle data and knee joint angle data at the target gait speed; and generating a gait phase feature vector at the target gait speed based on the initial gait phase using a gait rhythm generation model adapted to the target gait speed.

[0012] Furthermore, a nonlinear mapping model from gait phase features to knee joint motion features is constructed, including: constructing an initial nonlinear mapping model with gait phase feature vectors as input and knee joint angle trajectories as output; generating gait phase feature vectors at a reference walking speed according to the gait rhythm generation model, and obtaining knee joint angle data of the subject at the reference walking speed; training the initial nonlinear mapping model with the gait phase feature vectors at the reference walking speed and the knee joint angle data at the reference walking speed to obtain the nonlinear mapping model from gait phase features to knee joint motion features.

[0013] Furthermore, the initial nonlinear mapping model is a three-layer radial basis function neural network, including: an input layer, a nonlinear hidden layer, and an output layer; wherein, the nonlinear hidden layer uses Gaussian radial basis functions as activation functions, and the nonlinear hidden layer... The output of each hidden layer neuron is: , The input is the gait phase feature vector. For the first The center vector of each hidden layer neuron Let be the Euclidean norm of the input gait phase eigenvector and center vector. For the first The width parameter of each basis function; the output of the output layer is: , This represents the trajectory of the knee joint angle. This represents the total number of hidden layer neurons. For the first The connection weights from hidden layer neurons to the output layer This is the output layer bias term.

[0014] Furthermore, the initial nonlinear mapping model is trained using the gait phase feature vector at the reference gait speed and the knee joint angle data at the reference gait speed. This includes: inputting the gait phase feature vector at the reference gait speed into the initial nonlinear mapping model to obtain the predicted knee joint angle trajectory; and iteratively optimizing the parameters of the nonlinear hidden layer and the output layer with the training objective of minimizing the error between the predicted knee joint angle trajectory and the knee joint angle at the reference gait speed.

[0015] On the other hand, embodiments of the present invention provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described gait generation method for an embodied intelligent robot.

[0016] This invention provides a gait generation scheme that spans different walking speeds. First, a biomimetic continuous periodic rhythm generation central model capable of outputting gait phase feature vectors is constructed. Based on the subject's hip joint angle data at the target walking speed, the rhythmic oscillation angular frequency is adaptively adjusted to obtain a gait rhythm generation model adapted to the target walking speed and generate the corresponding gait phase feature vector. Furthermore, a nonlinear mapping model from gait phase features to knee joint motion features is established. The knee joint angle trajectory is output by the nonlinear mapping model to realize robot gait generation.

[0017] This invention employs a two-stage architecture combining rhythmic phase generation and nonlinear mapping. It retains the advantages of rhythmic generation models, such as rhythmic robustness, motion trajectory stability, biomimetic physiological rationality, and low computational overhead. At the same time, it leverages nonlinear mapping to achieve high-precision fitting of complex gait nonlinear characteristics. It can adapt to different target gait speeds by simply adjusting the rhythmic oscillation angular frequency. It can solve the knee joint angle at the target gait speed without retraining the model. This solves the problem that existing gait rhythm generation models are designed for fixed gait speeds and cannot achieve gait generation across gait speeds. It effectively improves the speed adaptation capability, cross-gait speed generalization performance, and engineering practicality of embodied intelligent robots.

[0018] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0019] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is an optional flowchart of a gait generation method for an embodied intelligent robot according to an embodiment of the present invention; Figure 2 This is an optional flowchart of a method for constructing a gait rhythm generation model according to an embodiment of the present invention; Figure 3 This is an optional flowchart of the rhythmic oscillation angular frequency calculation method according to an embodiment of the present invention; Figure 4This is an optional flowchart of a method for generating gait phase feature vectors at a target gait speed based on a gait rhythm generation model adapted to the target gait speed, according to an embodiment of the present invention. Figure 5 This is a schematic diagram comparing the predicted trajectory of the knee joint angle under the cross-velocity test with the measured gait in an embodiment of the present invention. Detailed Implementation

[0020] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0021] One specific embodiment of the present invention discloses a method for generating the gait of an embodied intelligent robot. Figure 1 This is an optional flowchart for a gait generation method for the embodied intelligent robot. For example... Figure 1 As shown, the method includes the following steps: S101: Constructing a gait rhythm generation model. This gait rhythm generation model is a biomimetic continuous-cycle rhythm generation central model used to generate gait phase feature vectors. As described in the background art, the biomimetic continuous-cycle rhythm generation central model simulates the CPG neural regulation mechanism of the motor center in vertebrates. By constructing an autonomous rhythmic oscillation system through a built-in coupled oscillatory neural network, it can autonomously generate stable, continuous, and periodic rhythmic oscillation signals. The biomimetic continuous-cycle rhythm generation central model can output continuous periodic gait phase feature vectors, which carry temporal information within the gait cycle and serve as a unified temporal reference for joint motion trajectory generation. The gait rhythm generation model can be constructed based on a reference gait speed. For example, the reference gait speed can be 1 m / s, but this invention is not limited thereto.

[0022] S102: Acquire hip joint angle data of the subject at the target gait speed. Adjust the rhythmic oscillation angular frequency of the gait rhythm generation model based on the hip joint angle data at the target gait speed to obtain a gait rhythm generation model adapted to the target gait speed. The hip joint motion cycle directly reflects the gait cycle. Solve for the rhythmic oscillation angular frequency based on the hip joint angle data and adjust the rhythmic oscillation angular frequency of the gait rhythm generation model to obtain the adjusted gait rhythm generation model. This ensures that the oscillation cycle of the gait rhythm generation model is consistent with the human gait cycle at the target gait speed, achieving matching between the gait rhythm generation model and the target gait speed, and completing the speed adaptive adjustment of the gait rhythm generation model. The target gait speed is the desired gait speed, which can be the same as or different from the reference gait speed.

[0023] S103: Generate a gait phase feature vector at the target gait speed based on the gait rhythm generation model adapted to the target gait speed. Specifically, the adjusted gait rhythm generation model continuously outputs a gait phase feature vector that changes continuously over time, providing standardized temporal input for subsequent solution of the knee joint angle trajectory.

[0024] S104: Construct a nonlinear mapping model from gait phase features to knee joint motion features. Input the gait phase feature vector at the target gait speed into the nonlinear mapping model, and output the knee joint angle trajectory at the target gait speed to achieve gait generation for the embodied intelligent robot. Establish a nonlinear mapping model with gait phase feature vector as input and knee joint angle trajectory as output. Input the gait phase feature vector corresponding to the target gait speed into the nonlinear mapping model, solve for the knee joint angle trajectory, and then use this knee joint angle trajectory to drive the joint movement of the embodied intelligent robot to complete the generation of the robot's biomimetic gait.

[0025] In the above embodiments, a gait generation scheme that spans different walking speeds is provided. First, a biomimetic continuous periodic rhythm generation central model capable of outputting gait phase feature vectors is constructed. Based on the subject's hip joint angle data at the target walking speed, the rhythmic oscillation angular frequency is adaptively adjusted to obtain a gait rhythm generation model adapted to the target walking speed and generate the corresponding gait phase feature vector. Furthermore, a nonlinear mapping model from gait phase features to knee joint motion features is established. The knee joint angle trajectory is output by the nonlinear mapping model to realize robot gait generation.

[0026] This invention employs a two-stage architecture combining rhythmic phase generation and nonlinear mapping. It retains the advantages of rhythmic generation models, such as rhythmic robustness, motion trajectory stability, biomimetic physiological rationality, and low computational overhead. Simultaneously, it leverages nonlinear mapping to achieve high-precision fitting of complex gait nonlinear characteristics. Only the rhythmic oscillation angular frequency needs to be adjusted to adapt to different target gait speeds, and the knee joint angle at the target gait speed can be solved without retraining the model. This solves the problem in existing gait rhythm generation models designed for fixed gait speeds, which cannot achieve gait generation across different gait speeds. It effectively improves the speed adaptation capability, cross-step speed generalization performance, and engineering practicality of embodied intelligent robots' gait. Furthermore, it addresses the problem in existing CPG gait generation technologies based on Hopf oscillators, where the limiting loop of the single circular Hopf oscillator is a standard circle, only capable of representing the basic periodic rhythm of gait. It cannot effectively fit the nonlinear gait characteristics caused by human interaction with the ground, impacts, and environmental noise during walking, and lacks the ability to express the details of complex joint movements, making it difficult to achieve high-precision gait trajectory reconstruction. It also solves the problem of simultaneously balancing the physiological rationality, rhythmic robustness, and nonlinear fitting accuracy of the CPG model, or the problem of introducing complex network structures to improve fitting accuracy, which leads to large computational overhead.

[0027] In a preferred embodiment of the present invention, a method for constructing a gait rhythm generation model is provided. Figure 2 An optional flowchart of the method is shown, such as Figure 2 As shown, it includes the following steps: S201: Obtain hip joint angle data of the subject at a reference walking speed; the hip joint angle data is time-domain data, that is, the collected hip joint angle time-series signal; S202: Calculate the rhythmic oscillation angular frequency of the gait rhythm generation model based on hip joint angle data at a reference walking speed; S203: Taking the minimum deviation between the gait phase feature vector generated by the gait rhythm generation model and the hip joint angle data under the reference walking speed as the optimization objective, determine the rhythm parameters of the gait rhythm generation model other than the rhythm oscillation angular frequency; S204: Based on other rhythm parameters and rhythm oscillation angular frequency, a biomimetic continuous periodic rhythm generation central model is constructed as a gait rhythm generation model under reference walking speed.

[0028] The rhythmic oscillation angular frequency is used to determine the gait oscillation period, while the remaining rhythmic parameters determine the oscillation limit loop morphology and phase evolution characteristics of the biomimetic continuous periodic rhythm generation central model. These two types of parameters have different functions, and a step-by-step identification strategy is adopted. First, the rhythmic oscillation angular frequency is extracted from the hip joint angle data. Then, this angular frequency is fixed, and the remaining rhythmic parameters are optimized to ensure that the gait phase timing output by the biomimetic continuous periodic rhythm generation central model matches the gait patterns of real humans, thus completing the construction of the gait rhythm generation model. Subsequent adaptation to different target gait speeds only requires adjusting the rhythmic oscillation angular frequency; the remaining rhythmic parameters do not need to be re-optimized. This provides support for the gait model's speed adaptation and cross-speed generalization capabilities, while ensuring that the gait rhythm generation model outputs a continuous and stable gait phase feature vector.

[0029] Regarding the angular frequency of rhythmic oscillations Figure 3 A flowchart illustrating an alternative method for calculating the angular frequency of rhythmic oscillations is shown, such as... Figure 3 As shown, it includes the following steps: S301: Perform a sliding window Fourier transform on the hip joint angle data to obtain the spectral information corresponding to each data window; S302: Calculate the instantaneous oscillation frequency of each data window based on the spectrum information; S303: Smooth the instantaneous oscillation frequency of each data window to obtain the smoothed oscillation frequency, and convert the smoothed oscillation frequency into the rhythmic oscillation angular frequency.

[0030] During the subject's walking process, the hip joint angle is a periodic temporal signal, and the fundamental frequency of the signal corresponds to the human gait cycle. Due to minor perturbations during walking and the influence of acquisition noise, the time-domain waveform is not a strictly constant periodic signal, making it difficult to capture the dynamic gait frequency using a single global Fourier transform. This method employs a sliding window Fourier transform to perform piecewise spectral analysis of the hip joint angle time sequence and calculate the rhythmic oscillation angular frequency. The specific steps include: 1) The collected hip joint angle data is divided into multiple overlapping data windows according to the preset window length and sliding step size. The angle data in each data window is processed by removing the mean, removing the trend and adding a window, and the corresponding spectrum is obtained by fast Fourier transform. 2) Within the preset gait frequency range, determine the fundamental frequency peak index of the m-th data window based on the spectral peak intensity and its harmonic relationship. ,according to Calculate the instantaneous oscillation frequency, where, Let N be the sampling frequency of the sliding window Fourier transform, and N be the number of periodic samples in the sliding window Fourier transform. The instantaneous oscillation frequency obtained by solving with continuous windows is susceptible to fluctuations due to motion jitter and sensor noise interference; therefore, the instantaneous oscillation frequency of the sequence corresponding to all windows is... Smoothing filtering is applied to suppress frequency abrupt changes and noise disturbances, and then the smoothed oscillation frequency is obtained after smoothing. A stable rhythmic oscillation angular frequency can be obtained through a conversion formula; 3) Based on the formula Smooth the oscillation frequency The rhythmic oscillation angular frequency required to transform into a biomimetic rhythm generation model .

[0031] In the above implementation, the hip joint angle data is analyzed in segments using a sliding window Fourier transform. Compared to the traditional global Fourier transform, this effectively adapts to the subtle dynamic changes in gait frequency during human walking. By limiting the effective frequency range of the gait and combining the relationship between spectral peak intensity and harmonics to accurately select the fundamental frequency peak, false frequency detections caused by clutter, higher harmonics, and acquisition noise are avoided, ensuring the accuracy of instantaneous gait frequency calculation. Simultaneously, the multi-window instantaneous oscillation frequency is smoothed and filtered to suppress interference, resulting in a stable, continuous, and realistic gait rhythm reference frequency. The final converted rhythmic oscillation angular frequency matches the subject's real gait cycle, providing a high-precision and highly stable rhythmic reference for subsequent gait rhythm adjustment at different target walking speeds, effectively improving the continuity of the robot's gait and the accuracy of speed adaptation.

[0032] Preferably, the biomimetic continuous periodic rhythm generation central model is a CPG model based on the Hopf oscillator, i.e., the Hopf-CPG model, which is as follows: ; in, x and y Let be the two two-dimensional state variables of the oscillator, and the right-hand side of the equation be the derivative of the gait phase eigenvector. The radial distance from the system state to the origin of the limit cycle is... ; is the convergence coefficient, used to adjust the speed at which the system state converges to the stable limit cycle; Let be the radius of the limiting cycle; The angular frequency of the rhythmic oscillation is used to determine the periodic characteristics of the oscillator; when When, the system state expands outward; when At that time, the system state contracts inward, and the system eventually converges to a radius of... A stable circular limit cycle.

[0033] Gait phase eigenvectors are composed of two-dimensional state variables and Together they form the expression: ; in, for The gait phase feature vector at time step is used to characterize the instantaneous phase position of the gait on the limiting cycle, reflecting the dynamic changes at different stages within the gait cycle. Based on the Hopf-CPG model, the output trajectory is always constrained by the convergence of the limiting cycle, greatly improving the safety of the output knee joint angle trajectory.

[0034] Based on the aforementioned Hopf-CPG model, this invention initializes the gait rhythm generation model by combining human hip and knee joint motion data during the gait phase feature vector generation stage at the target walking speed, thereby generating a gait phase feature vector adapted to the target speed. Figure 4 The flowchart illustrates an optional method for generating gait phase feature vectors at a target gait speed based on a gait rhythm generation model adapted to the target gait speed, such as... Figure 4 As shown, it specifically includes: S401: Acquire knee angle data of the subject at the target walking speed; S402: Construct the initial gait phase based on the hip joint angle data and knee joint angle data at the target gait speed; use the hip joint angle data at the target gait speed as a two-dimensional state variable. The knee joint angle data at the target walking speed is used as a two-dimensional state variable. Initial gait phase ; S403: Based on the initial gait phase, a gait rhythm generation model adapted to the target gait speed is used to generate the gait phase feature vector at the target gait speed. The initial gait phase is used as the initial condition of the Hopf-CPG model to calculate the gait phase feature vector at time t. The convergence coefficients in the Hopf-CPG model... and the radius of the limiting cycle Given (can be generated from presets), the angular frequency of rhythmic oscillation. The calculation yielded the gait phase feature vector at the target gait speed based on the Hopf-CPG model.

[0035] This invention utilizes hip and knee joint angle data to jointly construct the initial gait phase, replacing the random initialization method. This allows the Hopf oscillator in the Hopf-CPG model to directly operate in a stable limit cycle, suppressing transient oscillations during the startup phase and eliminating start-up jitter. At the same time, it ensures that the inherent phase coupling relationship of the bionic gait is maintained between the hip and knee joints, outputting a continuous and stable gait phase feature vector. This ensures that the gait does not jump during gait speed switching, improving the motion coordination and gait speed adaptive performance of the embodied intelligent robot. It is suitable for diverse walking scenarios of embodied intelligent robots.

[0036] In another preferred embodiment of the present invention, constructing a nonlinear mapping model from gait phase features to knee joint motion features includes: constructing an initial nonlinear mapping model with gait phase feature vectors as input and knee joint angle trajectories as output; generating gait phase feature vectors at a reference walking speed according to a gait rhythm generation model, and obtaining knee joint angle data of the subject at the reference walking speed; training the initial nonlinear mapping model with the gait phase feature vectors at the reference walking speed and the knee joint angle data at the reference walking speed to obtain a nonlinear mapping model from gait phase features to knee joint motion features.

[0037] Human knee joint angle data exhibits highly nonlinear and asymmetric motion patterns with gait phase, making it difficult for traditional linear mapping methods to accurately reproduce the trajectory of a realistic bionic knee joint angle. This implementation constructs a nonlinear mapping model that takes gait phase as input and knee joint angle trajectory as output. Training is completed using samples at a reference gait speed, allowing the model to fully learn the inherent nonlinear coupling relationship between gait phase and knee joint movement. Furthermore, model training is performed using only a single reference gait speed sample, and the model parameters are fixed after training, eliminating the need for re-collecting data and iterative training for different target gait speeds. When the embodied intelligent robot needs to switch between different gait speeds, only the rhythmic oscillation angular frequency of the Hopf-CPG model needs to be adjusted to generate a phase feature vector adapted to the target gait speed. Inputting this vector into the nonlinear mapping model directly yields the knee joint angle trajectory at the corresponding gait speed.

[0038] Therefore, this invention effectively overcomes the shortcomings of traditional gait generation models, such as the need for training with multiple velocity samples and poor cross-velocity adaptability. It significantly improves the cross-velocity generalization ability, adaptability, and engineering practicality of gait generation, while ensuring that the knee joint angle trajectory is smooth, highly biomimetic, and conforms to the real movement characteristics of the human body at each stride speed.

[0039] Optionally, the initial nonlinear mapping model is a three-layer radial basis function neural network (RBFNN), including: an input layer, a nonlinear hidden layer, and an output layer; wherein, the input layer is used to transfer the two-dimensional gait phase feature vector to the nonlinear hidden layer without transformation; the nonlinear hidden layer uses Gaussian radial basis functions as activation functions, and the nonlinear hidden layer... The output of each hidden layer neuron is: , The input is the gait phase feature vector. For the first The center vector of each hidden layer neuron Let be the Euclidean norm of the input gait phase eigenvector and center vector. For the first The width parameter of each basis function; the output of the output layer is: , This represents the trajectory of the knee joint angle. This represents the total number of hidden layer neurons. For the first The connection weights from hidden layer neurons to the output layer This is the output layer bias term.

[0040] In a specific implementation, for example, the input layer can be configured with 2 neurons, corresponding to the input two-dimensional gait phase feature vector, which are only responsible for transmitting the input features to the nonlinear hidden layer without transformation, without performing any additional computational processing; the nonlinear hidden layer is the core mapping unit of the network, and can be configured with a total of 50 hidden layer neurons, each neuron corresponding to a set of radial basis function center vectors and width parameters. This invention is not limited thereto.

[0041] This embodiment uses a Gaussian radial basis function as the activation function for the hidden layer neurons, and the width parameter of all hidden layer neurons can be uniformly set to 0.2 to complete the nonlinear transformation from a two-dimensional input space to a high-dimensional feature space. The linear output layer can be set to one neuron, corresponding to the angle prediction value of a single knee joint. The final knee joint angle prediction result is generated by linearly weighting and combining the outputs of all hidden layer neurons. This invention is not limited thereto.

[0042] This invention employs a three-layer radial basis function neural network to construct a phase-knee joint mapping model. This model is simple in structure, computationally inexpensive, and suitable for embedded real-time control scenarios in robots. The Gaussian radial basis function (RBF) possesses local activation characteristics, responding only to phase features near the basis function center. This allows for precise and detailed fitting of the complex, asymmetric, nonlinear coupling relationship between gait phase and knee joint movement, resulting in higher fitting accuracy and better trajectory smoothness compared to linear fitting methods. The output layer uses a linear weighted structure, enabling stable and rapid solution of continuous knee joint angle trajectories without complex iterative calculations. This RBFNN model, trained using only a single reference gait sample, can adapt to different gait phase features output by the CPG model, obtaining knee joint angle trajectories at different target gait speeds, thus enhancing the real-time performance and engineering practicality of gait generation for embodied intelligent robots.

[0043] The training process of the initial nonlinear mapping model includes: inputting the gait phase feature vector at a reference walking speed into the initial nonlinear mapping model to obtain the predicted knee joint angle trajectory; iteratively optimizing the parameters of the nonlinear hidden layer and output layer with the training objective of minimizing the error between the predicted knee joint angle trajectory and the knee joint angle at the reference walking speed. Using the gait phase feature vector as input to obtain the predicted knee joint angle trajectory, and using the synchronized real knee joint angle as the supervised target, the RBFNN nonlinear mapping model is trained, optimizing parameters such as the model's center vector, connection weights, and bias terms to effectively reduce the model's prediction error and ensure that the trained mapping model can accurately output a knee joint angle trajectory that conforms to the biomimetic characteristics of the human body. When the subject changes, the model parameters need to be retrained. Through a personalized retraining mechanism, the model can be adapted to the differences in lower limb movement among different subjects, improving the model's versatility and adaptability.

[0044] As an optional specific implementation, this invention uses gait sample data at a reference walking speed of 1.0 m / s to complete the training of the nonlinear mapping model. The specific process of model training is as follows: First, the lower limb kinematic data of two healthy subjects walking on flat ground at a reference walking speed of 1.0 m / s is collected using an optical motion capture system, with a sampling frequency of 100 Hz. The collected raw motion data is subjected to standardized preprocessing such as denoising, missing data completion, and gait cycle segmentation to accurately extract the temporal oscillation sequence of the hip and knee joints in the sagittal plane within a single complete gait cycle, constructing the real gait supervision dataset required for model training. Subsequently, a sliding window Fourier transform is performed on the preprocessed hip joint angle time-domain signal to extract the gait motion fundamental frequency information. Based on the fundamental frequency, the rhythmic oscillation angular frequency of the Hopf oscillator model is calculated and set to drive the rhythm generation model to run stably, outputting a continuous gait phase feature sequence that is completely aligned with the real human gait temporal sequence. Furthermore, this gait phase feature sequence is used as the sole input of the RBFNN, and the synchronously collected real knee joint angle data is used as supervision to complete the supervised training of the RBFNN nonlinear mapping model. This embodiment employs orthogonal least squares method to iteratively optimize the network parameters, obtaining the optimal combination of the RBFNN center vector, connection weights, and bias terms. Finally, the fitting accuracy is verified using a reference gait test set. Experimental results show that the trained model possesses extremely high fitting accuracy at the reference gait, and can fully and accurately learn the nonlinear flexion and extension motion of the knee joint at the human reference gait.

[0045] To comprehensively evaluate the cross-step speed generalization ability and speed adaptation performance of this invention, this embodiment adopts a cross-step speed verification strategy: RBFNN is trained offline using only a single reference step speed of 1.0 m / s. After training, all network parameters are fixed, and no further fine-tuning or retraining is performed for other speeds. During the cross-step speed test, the rhythmic oscillation angular frequency of the Hopf-CPG model is adaptively adjusted only based on the fundamental frequency of the hip joint motion signal at the target step speed to generate the gait phase feature sequence at the corresponding step speed, and then the knee joint angle trajectory at different target step speeds is inferred and output.

[0046] Experimental results are as follows Figure 5 As shown, the present invention maintains excellent trajectory prediction capabilities in scenarios with a slow walking speed of 0.8 m / s and a fast walking speed of 1.2 m / s. The determination coefficients for each test condition are all greater than 0.87, proving that the fixed-parameter RBFNN model can effectively adapt to different walking speeds and generate high-precision, cross-step speed knee joint angle trajectories without retraining.

[0047] like Figure 5As shown, the knee joint angle generated by this invention can completely cover the gait support phase range of 0° to 20° and the gait swing phase range of 0° to 70°, fully restoring the full-stage motion characteristics of a single-step gait. The average error between the generated knee joint angle trajectory and the measured knee joint angle trajectory is controlled within 10°, and the proportion of the support phase of the generated gait is highly consistent with the actual human gait, exhibiting excellent gait biomimicry and movement stability. Furthermore, experimental data from subjects show that the gait generated by this invention generally exhibits better fast-speed migration performance than slow-speed gait. While there are slight individual differences among different subjects, all can stably capture the core gait change patterns of different individuals, demonstrating good individual adaptability and versatility.

[0048] Compared with existing technologies, the gait generation scheme for embodied intelligent robots provided by this invention only requires a single offline training session based on a human gait sample with a single reference gait speed to completely solidify all network parameters of the nonlinear mapping model. During subsequent dynamic walking, when facing different target gait speed switching scenarios, there is no need to re-collect gait data or iteratively optimize and update network parameters. Simply by adjusting the rhythmic oscillation angular frequency of the Hopf-CPG model, it can adaptively match different walking speeds and output high-precision bionic knee joint trajectories. This overcomes the shortcomings of traditional gait generation models, such as reliance on multi-speed sample training, weak cross-speed generalization ability, and the need for repeated retraining and parameter tuning for speed changes. It significantly reduces online computational overhead and algorithm deployment costs, effectively improves the speed adaptation ability and cross-speed generalization performance of gait generation, and has strong practical application in engineering.

[0049] Another specific embodiment of the present invention discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described gait generation method for an embodied intelligent robot.

[0050] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0051] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating the gait of an embodied intelligent robot, characterized in that, include: A gait rhythm generation model is constructed; wherein, the gait rhythm generation model is a biomimetic continuous periodic rhythm generation central model, used to generate gait phase feature vectors; Acquire hip joint angle data of the subject at the target walking speed, and adjust the rhythmic oscillation angular frequency of the gait rhythm generation model according to the hip joint angle data at the target walking speed to obtain a gait rhythm generation model adapted to the target walking speed. Generate a gait phase feature vector at the target gait speed based on the gait rhythm generation model adapted to the target gait speed; A nonlinear mapping model is constructed from gait phase features to knee joint motion features. The gait phase feature vector at the target gait speed is input into the nonlinear mapping model, and the knee joint angle trajectory at the target gait speed is output to realize the gait generation of the embodied intelligent robot.

2. The gait generation method for an embodied intelligent robot according to claim 1, characterized in that, The construction of the gait rhythm generation model includes: Acquire the hip joint angle data of the subject at a reference walking speed; The rhythmic oscillation angular frequency of the gait rhythm generation model is calculated based on the hip joint angle data at the reference walking speed. With the goal of minimizing the deviation between the gait phase feature vector generated by the gait rhythm generation model and the hip joint angle data under the reference walking speed, other rhythm parameters of the gait rhythm generation model, excluding the rhythm oscillation angular frequency, are determined. Based on the other rhythm parameters and the rhythm oscillation angular frequency, the biomimetic continuous periodic rhythm generation central model is constructed as the gait rhythm generation model.

3. The gait generation method for an embodied intelligent robot according to claim 1 or 2, characterized in that, The rhythmic oscillation angular frequency is calculated as follows: Perform a sliding window Fourier transform on the hip joint angle data to obtain the spectral information corresponding to each data window; The instantaneous oscillation frequency of each data window is calculated based on the spectrum information; The instantaneous oscillation frequency of each data window is smoothed to obtain a smoothed oscillation frequency, and the smoothed oscillation frequency is converted into the rhythmic oscillation angular frequency.

4. The gait generation method for an embodied intelligent robot according to claim 3, characterized in that, The calculation of the instantaneous oscillation frequency of each data window based on the spectrum information includes: Based on the spectral peak intensity and harmonic relationship of each data window, determine the fundamental frequency peak index of the corresponding data window; Based on the fundamental frequency peak index and the sampling parameters of the sliding window Fourier transform, the instantaneous oscillation frequency of each data window is calculated.

5. The gait generation method for an embodied intelligent robot according to claim 4, characterized in that, The calculation of the instantaneous oscillation frequency of each data window based on the fundamental frequency peak index and the sampling parameters of the sliding window Fourier transform includes: The instantaneous oscillation frequency of each data window is calculated using the following formula: ； Where m is the sequence number of the data window, The instantaneous oscillation frequency is... This is the index of the fundamental frequency peak. The sampling frequency of the sliding window Fourier transform is denoted as N, and the number of periodic samples of the sliding window Fourier transform is denoted as N.

6. The gait generation method for an embodied intelligent robot according to claim 2, characterized in that, The gait phase feature vector at the target gait speed is generated based on the gait rhythm generation model adapted to the target gait speed, including: Acquire the knee joint angle data of the subject at the target walking speed; An initial gait phase is constructed based on the hip joint angle data and the knee joint angle data at the target gait speed. Based on the initial gait phase, the gait phase feature vector at the target gait speed is generated using the gait rhythm generation model adapted to the target gait speed.

7. The gait generation method for an embodied intelligent robot according to claim 1, characterized in that, The nonlinear mapping model for constructing gait phase features to knee joint motion features includes: Construct an initial nonlinear mapping model with the gait phase feature vector as input and the knee joint angle trajectory as output; The gait phase feature vector at a reference walking speed is generated based on the gait rhythm generation model, and the knee joint angle data of the subject at the reference walking speed is obtained. The initial nonlinear mapping model is trained using the gait phase feature vector at the reference gait speed and the knee joint angle data at the reference gait speed to obtain the nonlinear mapping model from the gait phase features to the knee joint motion features.

8. The gait generation method for an embodied intelligent robot according to claim 7, characterized in that, The initial nonlinear mapping model is a three-layer radial basis function neural network, including: an input layer, a nonlinear hidden layer, and an output layer; wherein, the nonlinear hidden layer uses Gaussian radial basis functions as activation functions, and the nonlinear hidden layer... The output of each hidden layer neuron is: , The input is the gait phase feature vector. For the first The center vector of each hidden layer neuron Let Euclidean norm be the input gait phase feature vector and the center vector. For the first The width parameter of each basis function; the output of the output layer is: , The knee joint angle trajectory is described. This represents the total number of neurons in the hidden layer. For the first The connection weights from hidden layer neurons to the output layer This is the output layer bias term.

9. The gait generation method for an embodied intelligent robot according to claim 8, characterized in that, The initial nonlinear mapping model is trained using the gait phase feature vector at the reference gait speed and the knee joint angle data at the reference gait speed, including: The gait phase feature vector at the reference walking speed is input into the initial nonlinear mapping model to obtain the predicted knee joint angle trajectory. With the training objective of minimizing the error between the predicted knee joint angle trajectory and the knee joint angle at the reference walking speed, the parameters of the nonlinear hidden layer and the output layer are iteratively optimized.

10. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the gait generation method for the embodied intelligent robot according to any one of claims 1 to 9.

Images