Voxel robot form optimization method based on hierarchical Bayesian probability model

Abstract

The invention discloses a voxel robot form optimization method based on a hierarchical Bayesian probability model, and the method comprises the steps: constructing a form generation model: constructing a hierarchical Bayesian probability generation model which comprises a task-individual-organ-voxel four-layer logic structure and is used for describing the relevance of tasks, individuals, organs and voxels; optimization of parameters of the generative model: estimating the parameters of the generative model by using the labeled data; performing form-control collaborative optimization: generating a robot form by using the generative model in a task environment, and performing fitness evaluation; screening a dominant form from fitness evaluation to update the generation model, and meanwhile, retaining a random generation form to maintain design diversity; introducing bionic priori: constructing a label-free natural biological form data set, and generating bionic priori through dimension reduction or feature learning; bionic prior and task label information are combined for optimization of the generative model. The method has high efficiency, interpretability and diversity.

Description

Technical Field

[0001] The present invention relates to the technical field of voxel robots, and in particular to a voxel robot morphology optimization method based on a hierarchical Bayesian probability model. Background Art

[0002] Voxel-based soft robots (VSRs) are made up of elastic modules (voxels) connected in a grid. Voxels can expand or contract under the combined action of external forces and their own actuators, thereby achieving robot posture changes. Compared with rigid robots made of rigid materials and joint actuators, VSRs have the advantages of flexibility, precision, and stability. They can not only complete simple tasks such as walking on flat ground, but also move freely on rugged terrain or complete complex object manipulation tasks. However, the complex structure of VSR also places higher demands on its morphological optimization process.

[0003] Most existing methods use a two-layer paradigm of "morphology-control collaborative optimization" to achieve VSR optimization. The inner control optimization uses various reinforcement learning algorithms to train control strategies for a given morphology to evaluate its adaptability to the environment. The outer morphology optimization iteratively updates the robot morphology based on fitness evaluation, and most of them are optimization algorithms. However, optimization algorithms tend to converge to a certain dominant morphology too early and cannot guarantee diversity.

[0004] In recent years, probabilistic generative models have been widely used in image generation and complex geometric structure modeling. Some studies have used probabilistic generative models to identify the dominant morphology of robots and maintain morphological diversity. However, existing probabilistic generative models use a simple "task-voxel" model structure, which results in large parameter scale and low optimization efficiency. Summary of the invention

[0005] In order to solve some or all of the technical problems existing in the above-mentioned prior art, the present invention provides a voxel robot morphology optimization method based on a hierarchical Bayesian probability model.

[0006] The technical solution of the present invention is as follows:

[0007] A voxel robot morphology optimization method based on a hierarchical Bayesian probability model is provided, the method comprising:

[0008] Constructing a morphological generation model: Constructing a hierarchical Bayesian probability generation model with a four-layer logical structure of "task-individual-organ-voxel" to describe the association between tasks, individuals, organs and voxels. The morphological generation process is described by modeling multiple distributions and their prior relationships between each layer;

[0009] Generative model parameter optimization: Use labeled data to estimate generative model parameters, introduce external prior information, and improve model generation efficiency and adaptability;

[0010] Morphology-control collaborative optimization: Generate robot morphology using generative models in task environments and perform fitness evaluation in combination with control strategy training; select dominant morphologies from fitness evaluation to update generative models, while retaining randomly generated morphologies to maintain design diversity;

[0011] Introducing bionic priors: constructing an unlabeled natural biological morphology dataset, generating bionic priors through dimensionality reduction or feature learning; combining bionic priors with task label information to optimize the generated model.

[0012] In one embodiment of the present invention, in the four-layer logical structure of "task-individual-organ-voxel" of the hierarchical Bayesian probability generation model: the distribution of each task on the individual obeys a multinomial distribution, and its parameters obey the Dirichlet prior; the distribution of each individual on the organ obeys a multinomial distribution, and its parameters obey the Dirichlet prior; the distribution of each organ on the voxel obeys a multinomial distribution, and its parameters obey the Dirichlet prior;

[0013] The morphology generation process of the generative model includes: extracting the individual type of the current robot morphology from the distribution of tasks to individuals, for each voxel of the robot morphology, extracting the organ type from the distribution of individuals to organs, and then extracting the voxel type from the distribution of organs to voxels.

[0014] In one embodiment of the present invention, in the optimization of the generative model parameters, stochastic variational inference technology is used to efficiently estimate the parameters of task distribution, individual distribution, organ distribution and voxel distribution, and to achieve the inference of latent variables.

[0015] In one embodiment of the present invention, in the generative model, position-related organ distribution parameters and voxel distribution parameters are used for voxels at different positions in the robot morphology to characterize the differentiated probability distribution of voxel positions.

[0016] In one embodiment of the present invention, the generative model improves the morphological generation process through variational autoencoder technology, including: using a neural network to generate organ and voxel type distributions of all voxels.

[0017] In one embodiment of the present invention, the morphology-control collaborative optimization is based on a generational update strategy, specifically including:

[0018] Generate several soft robot morphologies using generative models and place these morphologies in different task environments for control strategy training;

[0019] According to the fitness evaluation results, the dominant morphology is selected as the training sample with task label, and the parameters of the morphology generation model are updated;

[0020] In each generation of optimization, a certain proportion of randomly generated morphologies is retained to control the convergence rate and improve morphological diversity.

[0021] In one embodiment of the present invention, the morphology-control collaborative optimization is optimized with the evidence lower bound as the objective function, specifically including:

[0022] Through random variational inference, the joint distribution of individuals and organ types is modeled using latent variables z;

[0023] The optimization objective function includes the log-likelihood term logp of the generating process θ (x|z), the prior distribution term logp of the latent variable θ (z|h) and the prior distribution of the task label logp(h), and subtract the approximate posterior distribution The logarithmic probability of

[0024] The generative model parameters and the approximate form of the posterior distribution are adjusted according to the optimized lower bound of evidence to improve the adaptability and representation ability of the generative model.

[0025] In one embodiment of the present invention, the step of constructing an unlabeled natural biological morphology dataset and generating a bionic prior through dimensionality reduction or feature learning includes:

[0026] Projecting the unlabeled natural biological morphology dataset into the low-dimensional feature space of the morphology generation model to form a voxel matrix representation;

[0027] The distribution parameters of low-dimensional features are optimized through variational inference technology, from which morphological structures consistent with bionic features are extracted;

[0028] The generated biomimetic prior is used as an additional input to the morphology generation model to guide the model to converge to the biomimetic structural optimization.

[0029] In one embodiment of the present invention, the bionics prior and task label information are combined to generate model optimization, including:

[0030] Construct an optimization objective based on the evidence lower bound for unlabeled samples, including the log-likelihood term of the generation process and the prior distribution of the latent variables;

[0031] When optimizing the morphological generation model, labeled samples and unlabeled samples are combined at the same time, and the sample utilization efficiency and bionic structure adaptability of the generation model are improved through joint optimization.

[0032] The main advantages of the technical solution of the present invention are as follows:

[0033] The voxel robot morphology optimization method based on the hierarchical Bayesian probability model of the present invention effectively decomposes the high-dimensional complexity problem of voxel robot morphology design by constructing a hierarchical Bayesian probability generation model with a four-layer logical structure of "task-individual-organ-voxel", realizes accurate modeling of the relationship between tasks, morphological individuals and organs, and significantly improves the interpretability and efficiency of morphology generation. Through the optimization of the generation model parameters, the model can effectively utilize labeled data, quickly converge to the dominant morphology, and improve the task adaptability. The morphology-control collaborative optimization combines the control strategy with the morphology generation process, gradually selects the dominant morphology through fitness evaluation, and retains the randomly generated morphology, ensuring the diversity of design solutions and the ability to search for the global optimal solution. In addition, the introduction of bionic priors provides prior constraints for structural optimization through unlabeled natural biological morphology data, further improving the generation efficiency of the model and its adaptability to complex tasks. Overall, the method of the present invention has both high efficiency, interpretability and diversity, and provides an efficient and innovative solution for the automated design of voxel robot morphology. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0035] Figure 1 It is a flow chart of a voxel robot morphology optimization method based on a hierarchical Bayesian probability model according to an embodiment of the present invention;

[0036] Figure 2 It is a schematic diagram of the morphology generation process in the voxel robot morphology optimization method based on the hierarchical Bayesian probability model according to an embodiment of the present invention;

[0037] Figure 3 Schematic diagram of the approximate posterior process in the voxel robot morphology optimization method based on the hierarchical Bayesian probability model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0039] The technical solution provided by the embodiments of the present invention is described in detail below with reference to the accompanying drawings.

[0040] The mapping relationship between VSR morphology and environmental fitness has complex properties such as high dimension, non-convexity, and multi-peaks, which makes it difficult for traditional optimization algorithms such as genetic algorithms and Bayesian optimization to achieve ideal results. At the same time, the optimization algorithm is highly dependent on the initialization conditions, and it is easy to converge to the local optimum too early and it is difficult to find the optimal solution. In addition, the study found that for a given task, the optimization algorithm tends to converge to similar forms and it is difficult to ensure the diversity of design schemes. Although the existing probabilistic generation modeling scheme can effectively identify the dominant form and maintain the diversity of the design, it generally adopts the binary simple logic of "individual-voxel", which fails to effectively characterize the essential morphological characteristics of the dominant form and has low sample utilization efficiency. To this end, the present invention, inspired by the morphological muscle synergy theory, makes a four-layer logical assumption of "task-individual-organ-voxel" for the design problem to construct a VSR morphological generation model.

[0041] The embodiment of the present invention provides a voxel robot morphology optimization method based on a hierarchical Bayesian probability model, as shown in the attached Figure 1 As shown, including:

[0042] S1, constructing a morphological generation model: constructing a hierarchical Bayesian probability generation model with a four-layer logical structure of "task-individual-organ-voxel" to describe the association between tasks, individuals, organs and voxels. The morphological generation process is described by modeling each layer through multinomial distribution and its prior relationship.

[0043] In this embodiment, a hierarchical Bayesian probability generation model with a four-layer logical structure of "task-individual-organ-voxel" is first constructed to describe the association between tasks, individuals, organs and voxels. The association between each task type and the robot individual is modeled by multinomial distribution; the distribution between each individual type and the organ is modeled by multinomial distribution; the distribution between each organ and the voxel is modeled by multinomial distribution. In order to ensure the robustness and adaptability of the model, Dirichlet priors are introduced into each distribution parameter to constrain the distribution range of the parameters and enhance the generalization ability of the model.

[0044] By dividing the morphological generation model into four logical structures, describing the relationship between tasks, individuals, organs and voxels layer by layer, the complex morphological generation task is effectively decomposed, thereby reducing the search space of the model and improving the interpretability of the morphological generation process. This model not only promotes diversity through explicit division of "individual" types, but also, inspired by the theory of morphological physiology, innovatively proposes an "organ" hidden layer to characterize the synergistic relationship between voxels; by allowing the dominant morphologies of different tasks to share the "individual" and "organ" substructures, the problem scale is effectively simplified and the utilization rate of morphological samples is improved.

[0045] S2, Generative model parameter optimization: Use labeled data to estimate generative model parameters, introduce external prior information, and improve model generation efficiency and adaptability.

[0046] In this embodiment, labeled robot morphology samples are used to estimate and optimize the parameters of the morphology generation model, including the following steps: a training data set is given, including task labels and corresponding robot morphologies; the distribution parameters of each level in the morphology generation model (task→individual, individual→organ, organ→voxel) are estimated using the variational inference framework; the parameter optimization goal is to maximize the likelihood function of the labeled samples, and the model parameters are iteratively updated by back propagation until the model converges.

[0047] By using labeled data for supervised learning, the model can learn the structural characteristics of the dominant morphology under each task, thereby improving the morphology generation accuracy and efficiency of the generative model. Through parameter optimization, the correlation between modeling tasks and morphology is explicitly modeled, which improves the adaptability of the generative model in complex task environments; the use of labeled data accelerates the model training process and enables the model to converge efficiently.

[0048] S3, morphology-control collaborative optimization: Generate the robot morphology according to the morphology generation model in the task environment, and perform fitness evaluation in combination with control strategy training; select the dominant morphology from the fitness evaluation to update the generation model, while retaining the randomly generated morphology to maintain design diversity.

[0049] In the morphology-control collaborative optimization process, the system generates several soft robot morphologies according to the morphology generation model, including: using the morphology generation model to generate the initial robot morphology in the task environment; for each generated morphology, calculating the fitness of the morphology in the current task environment through control strategy training (such as control methods based on reinforcement learning); selecting the best performing dominant morphologies based on the fitness evaluation results, and using the task labels and morphology data corresponding to these morphologies to update the morphology generation model; retaining some randomly generated morphologies to prevent the model from converging prematurely, ensuring that the morphology design scheme has a certain diversity.

[0050] Morphology-control collaborative optimization is a training strategy based on generational updates. The training of morphology generation and control strategy cooperate with each other. The dominant morphology is selected through fitness evaluation and used for model update, thereby gradually improving the morphology design and task adaptability. Through the collaborative optimization of morphology generation and control strategy, the morphology design process can better adapt to the actual task environment; retaining the operation of randomly generating morphology helps maintain the diversity of morphology and avoid the model from falling into the local optimal solution.

[0051] S4, introduce bionic priors: construct an unlabeled natural biological morphology dataset, generate bionic priors through dimensionality reduction or feature learning; combine bionic priors with task label information to optimize the generated model.

[0052] In order to further improve the design efficiency and adaptability of the morphological generation model, an unlabeled natural biological morphological dataset is introduced in this embodiment, and a bionic prior is constructed through the following steps: morphological data of various organisms in nature are collected; feature extraction and dimensionality reduction are performed on these unlabeled data, and they are represented as a voxel matrix in a low-dimensional feature space; the generated bionic prior is input into the morphological generation model as an additional optimization constraint to help the model converge to a bionic structure.

[0053] Biological forms in nature have undergone long-term evolution and usually have highly adaptable structural characteristics. By introducing these unlabeled data, the model can obtain bionic priors during the training process, which helps the morphological design to optimize the dominant structure more efficiently. By introducing bionic priors, the design efficiency of the morphological generation model is improved; the use of unlabeled data reduces the cost of data collection and annotation, and expands the source of training data for the model.

[0054] In summary, the voxel robot morphology optimization method based on the hierarchical Bayesian probability model provided by the embodiment of the present invention effectively decomposes the high-dimensional complexity problem of voxel robot morphology design by constructing a hierarchical Bayesian probability generation model with a four-layer logical structure of "task-individual-organ-voxel", realizes accurate modeling of the relationship between tasks, morphological individuals and organs, and significantly improves the interpretability and efficiency of morphology generation. Through the optimization of the generation model parameters, the model can effectively utilize labeled data, quickly converge to the dominant morphology, and improve the task adaptability. The morphology-control collaborative optimization combines the control strategy with the morphology generation process, gradually selects the dominant morphology through fitness evaluation, and retains the randomly generated morphology, ensuring the diversity of design solutions and the ability to search for the global optimal solution. In addition, the introduction of bionics priors provides prior constraints for structural optimization through unlabeled natural biological morphology data, further improving the generation efficiency of the model and its adaptability to complex tasks. Overall, the method of the present invention has both high efficiency, interpretability and diversity, and provides an efficient and innovative solution for the automated design of voxel robot morphology, with broad application prospects.

[0055] The following is a detailed description of each step and the principles involved in the voxel robot morphology optimization method based on the hierarchical Bayesian probability model provided by an embodiment of the present invention.

[0056] 1. Morphological Generation Model

[0057] The present invention first expands the traditional morphological modeling idea of ​​"task-voxel" and proposes a Bayesian probability generation model with four-layer logic of "task-individual-organ-voxel". This model not only promotes diversity through explicit division of "individual" types, but also, inspired by the theory of morphological physiology, innovatively proposes the "organ" hidden layer to characterize the synergistic relationship between voxels; by allowing the dominant morphologies of different tasks to share the "individual" and "organ" substructures, it effectively simplifies the problem scale and improves the utilization rate of morphological samples.

[0058] Assume that there are H tasks, D soft robots, B robot types, K organs, and V voxels. The robot morphology consists of S a ×S b The two-dimensional voxel matrix representation, S a and S b denote the height and width of the robot respectively. Also assume that each task has a parameter g on B individuals h A multinomial distribution (h=1,…,H), each individual has parameter θ on K organs b A multinomial distribution (b=1,…,B) with parameters t=1,…,B for each organ on V types of voxels. multinomial distribution (k=1,…,K), and assume that the above distribution parameters all obey the Dirichlet prior: g h ~Dir(γ),θ b ~Dir(α), For a given task h, the robot shape d is generated as follows:

[0059] (1) From Y~multi(g h ) Extract the individual type Y of the current robot d ;

[0060] (2) For the nth voxel of the robot morphology:

[0061] a. By Type of organ extracted; b. The type of voxels to extract.

[0062] The present invention calls the above model a supervised hierarchical Bayesian probability generation model (Supervised Hierarchical Latent Dirichlet Allocation, S-HLDA), and uses stochastic variational inference (SVI) technology to perform parameter estimation and latent variable inference based on labeled dominant morphological samples.

[0063] To address the "bag-of-word" strong assumption problem in the generation process (under the condition of given distribution parameters, the generation process of each voxel is independent of each other, thus ignoring the relative position relationship between voxels), the present invention further proposes two solutions:

[0064] (1) Different organ distribution parameters are used for voxels at different positions in the robot form and voxel distribution parameters This enables differentiated probability distribution representation based on voxel positions.

[0065] (2) The variational auto-encoder (VAE) technology is used to characterize the morphological generation process, that is, the distribution of organs and voxel types of all voxels is generated at one time through a neural network, which not only improves the model representation ability, but also breaks the independence between different voxels and retains the ability to learn the position of data points. The present invention calls this model a supervised hierarchical variational auto-encoder (Supervised Hierarchical Variational Auto-Encoder, S-HVAE).

[0066] 2. Morphology-Control Collaborative Optimization

[0067] The present invention adopts the "morphology-control collaborative optimization" paradigm to train the above-mentioned morphology model. Specifically, an intergenerational update strategy is adopted, that is, each generation uses the morphology model to generate several robot morphologies, which are placed in different task environments for strategy training and fitness evaluation, and the dominant morphologies are selected as training samples with task labels for updating the morphology model; at the same time, a certain proportion of randomly generated morphologies are retained in each generation, so as to promote the morphology model to converge to the dominant morphology distribution while maximizing its learning ability for morphological diversity.

[0068] The optimization objective function is given below using the S-HVAE model as an example. The latent variables (i.e., individual and organ type) are abbreviated as z, the observable robot morphology and task type are denoted as x and h respectively, the generation process of the morphology model and the approximate posterior distribution are denoted as p and q respectively. For labeled samples, the evidence lower bound (ELBO) of the marginal likelihood function can be obtained according to formula (1) for variational inference:

[0069]

[0070]

[0071] in θ and They represent the estimated parameters of the generation process and the approximate posterior respectively.

[0072] 3. Utilization of unsupervised advantage

[0073] Many studies on morphological evolution computation have found that dominant morphologies usually converge to obvious bionic morphological features, such as a foot-like morphology with two legs and a reptile morphology that moves through a ventral foot, but few studies have attempted to summarize and utilize these natural priors. This paper innovatively constructs a natural biological morphology dataset without task labels, reduces its dimensionality to a voxel matrix, and uses variational inference technology to achieve unsupervised learning, thereby accelerating the convergence of the model to bionic structural optimization. For the above unlabeled samples, ELBO can be obtained according to formula (2):

[0074]

[0075] The present invention names the S-HVAE model after introducing unlabeled samples as Semi-Supervised Hierarchical Variational Auto-Encoder (SS-HVAE), and its optimization objective function is shown in formula (3), where and They represent the unlabeled and labeled sample sets respectively. The generation process is shown in the attached figure. Figure 2 The probabilistic graphical model shown in the figure shows that the approximate posterior inference is as shown in the attached figure. Figure 3 The probabilistic graphical model representation shown.

[0076]

[0077] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In addition, "front", "back", "left", "right", "upper" and "lower" in this article are all referenced to the placement state shown in the accompanying drawings.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A voxel robot morphology optimization method based on a hierarchical Bayesian probability model, characterized in that: include: Constructing a morphological generation model: Constructing a hierarchical Bayesian probability generation model with a four-layer logical structure of "task-individual-organ-voxel" to describe the association between tasks, individuals, organs and voxels. The morphological generation process is described by modeling multiple distributions and their prior relationships between each layer; Generative model parameter optimization: Use labeled data to estimate generative model parameters, introduce external prior information, and improve model generation efficiency and adaptability; Morphology-control collaborative optimization: Generate robot morphology using generative models in task environments and perform fitness evaluation in combination with control strategy training; select dominant morphologies from fitness evaluation to update generative models, while retaining randomly generated morphologies to maintain design diversity; Introducing bionic priors: constructing an unlabeled natural biological morphology dataset, generating bionic priors through dimensionality reduction or feature learning; combining bionic priors with task label information to optimize the generated model.

2. The voxel robot morphology optimization method based on hierarchical Bayesian probability model according to claim 1, characterized in that: In the four-layer logical structure of "task-individual-organ-voxel" of the hierarchical Bayesian probability generation model: the distribution of each task on the individual obeys a multinomial distribution, and its parameters obey the Dirichlet prior; the distribution of each individual on the organ obeys a multinomial distribution, and its parameters obey the Dirichlet prior; the distribution of each organ on the voxel obeys a multinomial distribution, and its parameters obey the Dirichlet prior; The morphology generation process of the generative model includes: extracting the individual type of the current robot morphology from the distribution of tasks to individuals, for each voxel of the robot morphology, extracting the organ type from the distribution of individuals to organs, and then extracting the voxel type from the distribution of organs to voxels.

3. The voxel robot morphology optimization method based on hierarchical Bayesian probability model according to claim 2, characterized in that: In the optimization of the generative model parameters, stochastic variational inference technology is used to efficiently estimate the parameters of task distribution, individual distribution, organ distribution and voxel distribution, and to achieve the inference of latent variables.

4. The voxel robot morphology optimization method based on hierarchical Bayesian probability model according to claim 3, characterized in that: In the generative model, position-related organ distribution parameters and voxel distribution parameters are used for voxels at different positions in the robot morphology to characterize the differentiated probability distribution of voxel positions.

5. The voxel robot morphology optimization method based on hierarchical Bayesian probability model according to claim 4, characterized in that: The generative model improves the morphological generation process through variational autoencoder technology, including: using a neural network to generate organ and voxel type distributions of all voxels.

6. The voxel robot morphology optimization method based on hierarchical Bayesian probability model according to claim 1, characterized in that: The morphology-control collaborative optimization is based on an intergenerational update strategy, specifically including: Generate several soft robot morphologies using generative models and place these morphologies in different task environments for control strategy training; According to the fitness evaluation results, the dominant morphology is selected as the training sample with task label, and the parameters of the morphology generation model are updated; In each generation of optimization, a certain proportion of randomly generated morphologies is retained to control the convergence rate and improve morphological diversity.

7. The voxel robot morphology optimization method based on hierarchical Bayesian probability model according to claim 6, characterized in that: The morphology-control collaborative optimization is optimized with the evidence lower bound as the objective function, specifically including: Through random variational inference, the joint distribution of individuals and organ types is modeled using latent variables z; The optimization objective function includes the log-likelihood term logp of the generating process θ (x|z), the prior distribution term logp of the latent variable θ (z|h) and the prior distribution of the task label logp(h), and subtract the approximate posterior distribution The logarithmic probability of The generative model parameters and the approximate form of the posterior distribution are adjusted according to the optimized lower bound of evidence to improve the adaptability and representation ability of the generative model.

8. The voxel robot morphology optimization method based on hierarchical Bayesian probability model according to claim 1, characterized in that: The method of constructing an unlabeled natural biological morphology dataset and generating a bionic prior through dimensionality reduction or feature learning includes: Projecting the unlabeled natural biological morphology dataset into the low-dimensional feature space of the morphology generation model to form a voxel matrix representation; The distribution parameters of low-dimensional features are optimized through variational inference technology, from which morphological structures consistent with bionic features are extracted; The generated biomimetic prior is used as an additional input to the morphology generation model to guide the model to converge to the biomimetic structural optimization.

9. The voxel robot morphology optimization method based on hierarchical Bayesian probability model according to claim 8, characterized in that: The bionic prior and task label information are combined to optimize the generation model, including: Construct an optimization objective based on the evidence lower bound for unlabeled samples, including the log-likelihood term of the generation process and the prior distribution of the latent variables; When optimizing the morphological generation model, labeled samples and unlabeled samples are combined at the same time, and the sample utilization efficiency and bionic structure adaptability of the generation model are improved through joint optimization.

Images