A method and system for generating dexterous hand motion trajectories based on variational autoencoders

Abstract

This invention provides a method and system for generating dexterous hand motion trajectories based on variational autoencoders, relating to the fields of robot motion control and artificial intelligence. This invention utilizes the encoding capability of a dexterous hand motion prior model to compress high-dimensional, complex temporal motion data into low-dimensional latent variables. Subsequently, the latent variables are input into the decoder of the dexterous hand motion prior model for decoding, thereby mapping the latent variables back to the high-dimensional motion space, realizing the reconstruction and generation of the target motion trajectory from the low-dimensional representation. Finally, the dexterous hand motion prior model ensures the accuracy of the generated trajectory through reconstruction error constraints, guaranteeing the correctness of the optimization direction of the dexterous hand motion prior model, and making the generated target motion trajectory approximate a realistic and reasonable motion trajectory. Ultimately, the generated target motion trajectory is naturally stable, conforms to the motion characteristics of a dexterous hand, and has good generalization ability.

Description

Technical Field

[0001] This invention relates to the fields of robot motion control and artificial intelligence technology, and in particular to a method, system, storage medium, robot system, and bitstream for generating the motion trajectory of a dexterous hand based on a variational autoencoder. Background Technology

[0002] Dexterous hand manipulation is an important research direction in robotics, involving high-dimensional joint collaboration, strong temporal dependence, and contact constraint coupling. Traditional rule-based or pure optimization methods are prone to problems such as unnatural trajectories, jitter, and low grasping success rates in complex scenarios.

[0003] Current technical challenges include: a lack of large-scale, high-quality, and multi-task-covering dexterous hand teaching data; highly diverse and strongly coupled dexterous hand movements; significant imbalances in movement distribution; and extremely high requirements for temporal continuity and physical feasibility in dexterous hand trajectories.

[0004] Although the field of human motion modeling has proposed hand motion priors for human hands, the modeling object of hand motion priors is usually the parameter space of the human hand, which differs significantly from the dexterous hand of a robot in terms of joint structure, degree of freedom configuration, range of motion, joint coupling relationship and execution constraints. Directly using human hand priors often makes it difficult to guarantee the executability of robot motion and task consistency. Summary of the Invention

[0005] In view of this, the core technical problem to be solved by the present invention is: how to provide a method that can generate a natural, stable motion trajectory that conforms to the characteristics of dexterous hand movement.

[0006] On one hand, the present invention provides a method for generating dexterous hand motion trajectories based on variational autoencoders, comprising the following steps: S1, acquiring preprocessed dexterous hand motion time-series data; S2, inputting the dexterous hand motion time-series data into a trained dexterous hand motion prior model to obtain latent variables; S3, inputting the latent variables into the decoder of the dexterous hand motion prior model for decoding to generate the target motion trajectory of the dexterous hand; wherein, the dexterous hand motion prior model is a generative neural network model used to learn the prior distribution of dexterous hand motion, and the dexterous hand motion prior model is trained by minimizing the weighted combination of reconstruction error and latent variable distribution regularization term as the objective function, and the reconstruction error is the error between the target motion trajectory and the actual motion trajectory of the dexterous hand.

[0007] Optionally, step S1 specifically includes the following steps: S11, acquiring human hand motion capture data; S12, mapping the human hand motion capture data to the motion space of the target dexterous hand using a motion mapping algorithm to obtain the original dexterous hand joint angle sequence; S13, preprocessing the original dexterous hand joint angle sequence to obtain preprocessed dexterous hand motion time series data; wherein, the dexterous hand motion time series data includes the joint angles and joint angular velocities of the dexterous hand, and the preprocessing includes at least one of time alignment, sequence segmentation, scale normalization, and statistical standardization.

[0008] Optionally, in step S12, the motion mapping algorithm is a gradient-based inverse kinematics optimization algorithm based on vector constraints.

[0009] Optionally, in step S3, the encoder of the dexterous hand motion prior model is a temporal encoder containing a residual structure.

[0010] Optionally, in step S3, the decoder of the dexterous hand motion prior model is a decoder based on the neural motion field.

[0011] Optionally, in step S3, the reconstruction error includes joint angle error and joint angular velocity error.

[0012] Optionally, the dexterous hand motion trajectory generation method is applied to dexterous hand imitation learning, trajectory planning or motion control tasks, and the generated target motion trajectory is used to drive the dexterous hand to perform operation tasks.

[0013] On one hand, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the dexterity hand motion trajectory generation method described above.

[0014] On one hand, the present invention also provides a robot system, including: a processor, a memory, and a device communicatively connected to the processor. The memory stores a computer program, which, when executed by the processor, performs the dexterous hand motion trajectory generation method as described above and generates a target motion trajectory for driving the dexterous hand to perform an operation task; and a dexterous hand, communicatively connected to the memory and the processor, for receiving the target motion trajectory and performing corresponding actions.

[0015] On the other hand, this invention provides a dexterity hand motion trajectory generation system based on variational autoencoders, used to implement the dexterity hand motion trajectory generation method described above, comprising: a data acquisition module, used to acquire preprocessed dexterity hand motion time-series data, the dexterity hand motion time-series data including joint angles and joint angular velocities of the dexterity hand; a latent variable generation module, communicatively connected to the data acquisition module, used to input the dexterity hand motion time-series data into a trained dexterity hand motion prior model to obtain latent variables, the dexterity hand motion prior model being a generative neural network model used to learn the prior distribution of dexterity hand motion; and a trajectory generation module, communicatively connected to the latent variable generation module and the data acquisition module, used to input the latent variables into the decoder of the dexterity hand motion prior model for decoding to generate the target motion trajectory of the dexterity hand; wherein, the dexterity hand motion prior model is trained by minimizing a weighted combination of reconstruction error and latent variable distribution regularization term as the objective function, the reconstruction error being the error between the target motion trajectory and the actual motion trajectory of the dexterity hand.

[0016] On the other hand, the present invention provides a bitstream containing target motion trajectory data generated by the above-described dexterous hand motion trajectory generation method, wherein the bitstream is used to represent a series of continuous joint angle and / or joint angular velocity commands of the dexterous hand.

[0017] On the other hand, the present invention also provides a method for storing a bit stream, comprising: performing the above-described method for generating a dexterous hand motion trajectory to generate a target motion trajectory of a dexterous hand; and storing the bit stream representing the target motion trajectory into a storage medium.

[0018] The implementation of this invention offers the following advantages: By employing a dexterous hand motion prior model based on a variational autoencoder, this invention can learn the prior distribution of dexterous hand motion from large-scale teaching data, generating natural and stable trajectories that conform to the characteristics of dexterous hand motion. Specifically, through the encoding capability of the dexterous hand motion prior model, high-dimensional and complex temporal motion data is compressed into low-dimensional latent variables, realizing the extraction and representation of the essential features of dexterous hand motion, laying the foundation for generating motion trajectories that conform to the prior distribution. Subsequently, the latent variables are input into the decoder of the dexterous hand motion prior model for decoding, thereby mapping the latent variables back to the high-dimensional motion space, realizing the reconstruction and generation of the target motion trajectory from the low-dimensional representation. Simultaneously, the dexterous hand motion prior model adopts a variational autoencoder architecture, whose inherent encoder-decoder structure and regularization reduction of the latent space distribution... The constraints enable the dexterous hand motion prior model to learn a structured, continuous, and sampleable dexterous hand motion prior distribution, rather than a simple data mapping. Finally, the dexterous hand motion prior model ensures the accuracy of the generated trajectory through reconstruction error constraints and the rationality of the latent variable distribution through distribution regularization constraints, preventing overfitting or the generation of meaningless latent variables. The reconstruction error is the error between the target motion trajectory and the actual motion trajectory of the dexterous hand, ensuring the correctness of the optimization direction of the dexterous hand motion prior model, making the generated target motion trajectory approximate the real and reasonable motion trajectory. These technical features work synergistically to ultimately achieve a naturally stable target motion trajectory that conforms to the motion characteristics of a dexterous hand and has good generalization ability, solving problems such as high-frequency jitter, abrupt posture changes, contact discontinuities, and inconsistencies with hardware constraints that are easily encountered when generating trajectories in high-dimensional joint space using traditional methods. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for generating motion trajectories of a dexterous hand based on a variational autoencoder, as shown in one embodiment. Figure 2 This is the overall architecture of a dexterous hand motion prior model in one embodiment; Figure 3 As one embodiment, the encoder architecture of a dexterous hand motion prior model; Figure 4 In one embodiment, a visual representation of the motion prior reconstruction results of a dexterous hand with fine manipulation is provided. Figure 5 In one embodiment, a visual representation of the motion prior reconstruction results of a dexterous hand under perturbation conditions; Figure 6 In one embodiment, the motion prior of dexterous hand fine manipulation is used for downstream dexterous hand imitation learning demonstration. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0021] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the embodiments of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0022] In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims. In the description of this application, it should be understood that the terms "first," "second," "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0023] Furthermore, in the description of this application, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0024] Example 1 In this embodiment, as Figure 1 The method for generating the motion trajectory of a dexterous hand based on a variational autoencoder, as shown, includes the following steps: S1. Obtain the preprocessed dexterous hand motion timing data; S2. Input the dexterity hand motion time series data into the trained dexterity hand motion prior (DMP) model to obtain latent variables; S3. Input the latent variables into the decoder of the dexterous hand motion prior model for decoding to generate the target motion trajectory of the dexterous hand; Among them, the dexterity hand motion prior model is a generative neural network model used to learn the prior distribution of dexterity hand motion. The dexterity hand motion prior model is trained by minimizing the weighted combination of reconstruction error and latent variable distribution regularization term as the objective function. The reconstruction error is the error between the target motion trajectory and the actual motion trajectory of the dexterity hand.

[0025] In this embodiment, the dexterous hand motion trajectory generation method is applied to dexterous hand imitation learning, trajectory planning or motion control tasks, and the generated target motion trajectory is used to drive the dexterous hand to perform operation tasks.

[0026] In this embodiment, step S1 specifically includes the following steps: S11. Acquire human hand motion capture data; S12. The motion capture data of the human hand is mapped to the motion space of the target dexterous hand through a motion mapping algorithm to obtain the original dexterous hand joint angle sequence. S13. Preprocess the original dexterous hand joint angle sequence to obtain preprocessed dexterous hand motion time series data; wherein, the dexterous hand motion time series data includes the joint angles and joint angular velocities of the dexterous hand, and the preprocessing includes at least one of time alignment, sequence segmentation, scale normalization and statistical standardization.

[0027] In this embodiment, the time-series data of dexterous hand movements are processed by time alignment, sequence segmentation, scale normalization, and statistical standardization to balance data distribution, reduce noise impact, and improve network training stability and convergence speed.

[0028] In this embodiment, the teaching data can be sourced from GRAB (GRasping Actions with Bodies).

[0029] Specifically, such as Figure 2 As shown, this embodiment uses datasets such as GRAB as the source of human hand motion capture trajectories. Datasets such as GRAB provide hand joint pose sequences and their interaction information with objects based on the MANO model (Modeled Hands, a parametric hand model), which can reflect various typical operational actions and their temporal characteristics.

[0030] In this embodiment, in step S12, the motion mapping algorithm is a gradient-type inverse kinematics optimization algorithm based on vector constraints, which solves the joint configuration of the dexterous hand by minimizing the difference between the vectors between key points of the human hand and the vectors between the corresponding key points of the dexterous hand.

[0031] Specifically, the MANO (hand Model with Articulated and Non-rigid defOrmations) human hand motion sequence is mapped to the motion space of the target dexterous hand through a redirection algorithm to obtain joint angular trajectories and joint angular velocity trajectories consistent with the degrees of freedom of the dexterous hand.

[0032] Specifically, since the MANO hand and the target dexterous hand differ in terms of joint topology, degree of freedom configuration, and joint range, this embodiment uses a gradient-based inverse kinematics optimization framework based on vector constraints to map the MANO key point sequence to the dexterous hand joint space.

[0033] Specifically, at each frame time t, the MANO joint is first transformed to a unified VR coordinate system and a local wrist reference system to obtain the positions of key points on the human body. Select human point pairs based on index pairs and construct target vectors. The specific formula is as follows: Formula (1); in, , Represents a specified human hand joint (e.g., finger bone segments, wrist base, and fingertips). Target vector It represents a vector formed by the finger bone segments, the base of the wrist, and the fingertips of the five fingers, in order to enhance the ability to express the fine posture of the fingertips and the intention of the end-effector.

[0034] On the dexterous hand side, the corresponding vector is calculated using forward kinematics. Let... For the joint angles of the current dexterous hand, , Let represent the three-dimensional positions of the starting joint and the target joint, respectively. Then the dexterous hand vector is: Formula (2); The retargeting process solves for the joint angles of each frame by minimizing the following objective function: Formula (3); in, The loss between the predicted result and the true value is calculated; s is a scale factor constant, which is adaptively set according to the hand size of the dexterous hand and the MANO model. s is set to 1.2 to compensate for the size difference between the human hand and the dexterous hand; λ is the smoothing regularization weight, typically set to 0.004 to ensure continuous joint angle changes between adjacent frames and reduce jitter. Through the above redirection steps, the dexterous hand joint angle sequence used for training the dexterous hand motion prior model is obtained. and its joint angular velocity sequence .

[0035] In this embodiment, the dexterous hand motion prior model is specifically a network model based on a variational autoencoder.

[0036] Specifically, training the dexterous hand motion prior model includes constructing and training a motion prior encoding network and constructing and training a motion decoding generation network.

[0037] Specifically, constructing and training the motion prior encoding network includes: Using preprocessed dexterous hand joint angles and angular velocities as input, feature extraction and motion modeling are first performed via a residual temporal encoder. This invention introduces multi-layer residual modules into the encoder of the dexterous hand motion prior model to enhance temporal multi-scale feature extraction capabilities and alleviate gradient vanishing and feature degradation problems in deep network training. The encoder output of the dexterous hand motion prior model is used to parameterize the latent variable distribution, generating the mean and variance of the latent space representation, thus compressing the dexterous hand temporal data into a compact latent representation. Kullback-Leibler Divergence (KL divergence) constraint aligns the latent space with the prior distribution, thereby obtaining a sampleable dexterous hand motion prior distribution.

[0038] Specifically, constructing and training the motion decoding generation network includes: In the decoding phase (where latent variables are input into the decoder of the dexterity hand motion prior model), the latent variables are used as conditional inputs, and a motion decoder based on Neural Motion Fields (NeMF) is used to reconstruct the dexterity hand motion in continuous time. The decoder outputs the joint angle trajectory at any time point with continuous time as the independent variable, achieving continuous, differentiable, and high-fidelity reconstruction of the entire motion. Compared to discrete frame-by-frame regression, latent neural fields can more naturally characterize the smoothness and local details of the motion trajectory. During training, the reconstruction error of joint angles and joint angular velocities is used as the main objective function, and latent space regularization and temporal smoothing terms are jointly optimized to finally obtain a dexterity hand motion prior model that can characterize the distribution of dexterity hand movements.

[0039] In this embodiment, as Figure 2 As shown, this embodiment performs preprocessing on the temporal characteristics of dexterous hand movements, as detailed below: The joint angles and joint angular velocities are standardized and normalized. Time alignment and segmentation strategies are used for sequences of different lengths to form fixed time window segments.

[0040] In this embodiment, as Figure 2As shown, this embodiment uses a variational autoencoder (VAE) to model the motion priors of a dexterous hand. The VAE includes an encoder and a decoder based on Neural Motion Fields (NeMF).

[0041] In this embodiment, the VAE encoder is a timing encoder that includes a residual structure, which measures the joint angle. With joint angular velocity The sequence is concatenated into input x and then input into a timing encoder. The formula for calculating input x is as follows: Formula (4).

[0042] In this embodiment, as Figure 3 As shown, the encoder's residual structure contains multiple layers of custom residual blocks, each containing a main branch and a shortcut connection branch.

[0043] In one specific embodiment, the main branch uses a skeleton convolution with a kernel size of 4 and a stride of 2 to extract spatiotemporal features, combined with grouping normalization and the Tanh activation function; the shortcut branch uses a skeleton convolution with a kernel size of 1 and a stride of 2 to complete channel mapping and temporal downsampling. After the two branches are added together, adjacent joint information is gradually aggregated through average pooling. As the number of network layers increases, the number of feature channels increases layer by layer, while the temporal resolution decreases layer by layer. Finally, the extracted features are flattened, and the mean and logarithmic variance of the 1024-dimensional latent variables are output through two independent fully connected layers. The encoder output is used to parameterize the latent variable distribution. The specific calculation formula is as follows: Formula (5); Using reparameterization techniques to analyze latent variables The sampling formula is as follows: Formula (6); in .

[0044] In this embodiment, the training objective of VAE Defined as reconstruction error The loss function, weighted by the regularization term of the Kullback-Leibler Divergence constraint, is as follows: Formula (7); in β is the weight, and its value is set to 1e based on empirical values. -5 Reconstruction error can be defined as a loss function of a weighted combination of joint angle error and angular velocity error, as follows: Formula (8); in, To reconstruct the joint angle and joint angular velocity output by the decoder respectively.

[0045] In this embodiment, a NeMF-based decoder is used to reconstruct the dexterous hand motion in continuous time. The decoder first maps time *t* to periodic features using Fourier position encoding across seven frequency bands, and concatenates this with a 1024-dimensional latent variable *z*. Then, it inputs a network consisting of eight layers of an MLP (Multi-Layer Perceptron) with a hidden dimension of 1024. Each layer employs LayerNorm (Layer Normalization), ReLU (Rectified Linear Unit), and skip connections to ultimately predict the joint angles and angular velocities at a given moment. This continuous-time representation allows for querying the motion state at any point in time, enabling smooth motion generation independent of the sampling frame rate. The NeMF decoder takes time *t* and the latent variable *z* as input and outputs the joint angles at that moment. : Formula (9); This allows NeMF to represent the entire motion trajectory using an implicit neural field. Compared to discrete frame-by-frame regression, NeMF, with its continuous-time modeling, can more naturally express trajectory smoothness and local details, and has a stronger adaptability to sequences of different lengths and sampling rates.

[0046] To verify the reconstruction accuracy of the dexterity hand motion prior model, given the lack of a unified evaluation standard for dexterity hand motion reconstruction in existing studies, this embodiment proposes and adopts the following index system, and verifies it on the test set.

[0047] Joint Angle Error : Formula (10); Used to quantify attitude reconstruction accuracy.

[0048] Joint Velocity Error : Formula (11); Used to quantify the degree of fit between dynamic consistency and temporal changes.

[0049] Key point spatial error : The key points are obtained from forward kinematics. Let... The key point locations are obtained from the actual joint angles. To reconstruct the key point locations, then: Formula (12); Key points can include the fingertips, the ends of the knuckles, and other operation-related link positions. The key point spatial error index maps joint spatial errors to more physically meaningful three-dimensional geometric deviations.

[0050] The above metrics were statistically analyzed on the test set, and the resulting quantitative results of the reconstruction accuracy and stability of the dexterity hand motion prior model on unseen sequences can be used to verify the effectiveness and generalization ability of the dexterity hand motion prior model. The final results are as follows: Figure 4 and Figure 5 As shown: Figure 4 The study demonstrates that the motion trajectory output by the dexterous hand motion prior model is highly consistent with the motion posture of the dexterous hand in real-world scenarios. Validation on a test set shows that the trajectory generated by this method has an average joint angle error of less than 1.6 degrees and an average keypoint spatial error of less than 2.4 millimeters, proving its high accuracy and effectiveness. This indicates that the dexterous hand motion prior model can achieve high-precision hand motion reconstruction. Figure 5 Even when the dexterous hand joints are disturbed by noise, the dexterous hand motion prior model can still use the learned motions to reconstruct reasonable dexterous hand movements.

[0051] In this embodiment, Figure 6 It demonstrates that downstream imitation learning strategies, guided by a dexterous hand motion prior model, can accomplish fine manipulation tasks.

[0052] Example 2 Based on Embodiment 1, this embodiment provides a dexterous hand motion trajectory generation system for implementing the dexterous hand motion trajectory generation method of Embodiment 1.

[0053] In this embodiment, the dexterous hand motion trajectory generation system based on variational autoencoder includes: The data acquisition module is used to acquire preprocessed dexterity hand motion time series data, which includes the joint angles and joint angular velocities of the dexterity hand. The latent variable generation module is connected to the data acquisition module. The latent variable generation module is used to input time series data into the trained dexterity hand motion prior model to obtain latent variables. The dexterity hand motion prior model is a generative neural network model used to learn the prior distribution of dexterity hand motion. The trajectory generation module is connected to the latent variable generation module and the data acquisition module. It is used to input the latent variables into the decoder of the dexterous hand motion prior model for decoding and to generate the target motion trajectory of the dexterous hand. Among them, the dexterous hand motion prior model is trained by minimizing the weighted combination of reconstruction error and latent variable distribution regularization term as the objective function, and the reconstruction error is the error between the target motion trajectory and the actual motion trajectory of the dexterous hand.

[0054] Example 3 This embodiment provides a computer-readable storage medium based on the above embodiments.

[0055] In this embodiment, a computer-readable storage medium stores computer instructions that are used to cause a computer to execute the dexterity hand motion trajectory generation method in Embodiment 1.

[0056] Example 4 This embodiment provides a robot system based on the above embodiments.

[0057] In this embodiment, the robot system includes: processor, The memory is connected to the processor and stores a computer program. When the computer program is executed by the processor, it executes the dexterous hand motion trajectory generation method as described in Embodiment 1 and generates a target motion trajectory for driving the dexterous hand to perform an operation task. The dexterous hand is connected to a memory and processor to receive the target's motion trajectory and execute corresponding actions.

[0058] Example 5 Based on the above embodiments, this embodiment provides a bitstream containing target motion trajectory data generated by the dexterous hand motion trajectory generation method of Embodiment 1.

[0059] In this embodiment, the bit stream is used to represent a series of continuous joint angle and / or joint angular velocity commands of the dexterous hand.

[0060] Example 6 This embodiment provides a method for storing bit streams, based on the above embodiments.

[0061] In this embodiment, the method for storing the bit stream includes: generating the target motion trajectory of the dexterous hand by executing the dexterous hand motion trajectory generation method of Embodiment 1; and storing the bit stream representing the target motion trajectory into a storage medium.

[0062] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for generating the motion trajectory of a dexterous hand based on a variational autoencoder, characterized in that, Includes the following steps: S1. Obtain the preprocessed dexterous hand motion timing data; S2. Input the dexterous hand motion time series data into the trained dexterous hand motion prior model to obtain latent variables; S3. Input the latent variables into the decoder of the dexterous hand motion prior model for decoding to generate the target motion trajectory of the dexterous hand; The dexterity hand motion prior model is a generative neural network model used to learn the prior distribution of dexterity hand motion. The dexterity hand motion prior model is trained by minimizing the weighted combination of reconstruction error and latent variable distribution regularization term as the objective function. The reconstruction error is the error between the target motion trajectory and the actual motion trajectory of the dexterity hand. The reconstruction error includes joint angle error and joint angular velocity error. The dexterity hand motion prior model ensures the accuracy of the generated trajectory through reconstruction error constraints and ensures the rationality of the latent variable distribution through distribution regularization term constraints, preventing the model from overfitting or generating meaningless latent variables. The encoder of the dexterous hand motion prior model is a temporal encoder containing a residual structure; The decoder for the dexterous hand motion prior model is a decoder based on the neural motion field.

2. The method for generating the motion trajectory of a dexterous hand as described in claim 1, characterized in that, Step S1 specifically includes the following steps: S11. Acquire human hand motion capture data; S12. The human hand motion capture data is mapped to the motion space of the target dexterous hand using a motion mapping algorithm to obtain the original dexterous hand joint angle sequence; S13. Preprocess the original dexterous hand joint angle sequence to obtain the preprocessed dexterous hand motion timing data; The dexterity hand motion time series data includes the joint angles and joint angular velocities of the dexterity hand, and the preprocessing includes at least one of time alignment, sequence segmentation, scale normalization, and statistical normalization.

3. The method for generating the motion trajectory of a dexterous hand as described in claim 2, characterized in that, In step S12, the motion mapping algorithm is a gradient-type inverse kinematics optimization algorithm based on vector constraints.

4. The method for generating the motion trajectory of a dexterous hand as described in claim 1, characterized in that, The method for generating the motion trajectory of a dexterous hand is applied to tasks such as imitation learning, trajectory planning, or motion control of a dexterous hand. The generated target motion trajectory is used to drive the dexterous hand to perform operational tasks.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to perform the dexterity hand motion trajectory generation method as described in any one of claims 1 to 4.

6. A robot system, characterized in that, include: processor, The memory is communicatively connected to the processor. The memory stores a computer program that, when executed by the processor, performs the dexterity hand motion trajectory generation method as described in any one of claims 1 to 4, and generates a target motion trajectory for driving the dexterity hand to perform an operation task. The dexterous hand is communicatively connected to the memory and the processor, and is used to receive the target motion trajectory and perform corresponding actions.

7. A dexterous hand motion trajectory generation system based on a variational autoencoder, used to implement the dexterous hand motion trajectory generation method as described in any one of claims 1 to 4, characterized in that, include: The data acquisition module is used to acquire preprocessed dexterity hand motion timing data, which includes the joint angles and joint angular velocities of the dexterity hand. The latent variable generation module is communicatively connected to the data acquisition module. The latent variable generation module is used to input the time series data of the dexterous hand movement into the trained dexterous hand movement prior model to obtain latent variables. The dexterous hand movement prior model is a generative neural network model used to learn the prior distribution of dexterous hand movement. The trajectory generation module is communicatively connected to the latent variable generation module and the data acquisition module, and is used to input the latent variables into the decoder of the dexterous hand motion prior model for decoding, so as to generate the target motion trajectory of the dexterous hand; The dexterity hand motion prior model is trained by minimizing the weighted combination of reconstruction error and latent variable distribution regularization term as the objective function, where the reconstruction error is the error between the target motion trajectory and the actual motion trajectory of the dexterity hand.

8. A bitstream comprising target motion trajectory data generated by the dexterous hand motion trajectory generation method of claim 1, characterized in that, The bitstream is used to represent a series of continuous joint angle and / or joint angular velocity commands for a dexterous hand.

9. A method for storing a bit stream, characterized in that, include: The method for generating the motion trajectory of a dexterous hand as described in claim 1 is used to generate the target motion trajectory of the dexterous hand; And store the bit stream representing the target's motion trajectory to a storage medium.

Images