Morphological statistical modeling-based voxel robot structure optimization design method
Through the depth probability generation model and morphological pooling technology, the morphological design of voxel robots is optimized, and the problem of relying on expert experience and limited multitasking adaptability in the existing technology is solved, and efficient morphological design optimization and multitasking adaptability are achieved.
Patent Information
- Application Number
- CN202411964100.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-30
- Publication Date
- 2025-05-16
AI Technical Summary
In the optimization of voxel robot morphological design, the prior art has problems such as relying on expert experience, limited adaptability to multi-task scenarios, and large morphological search space, making it difficult to efficiently obtain the optimal solution.
Using a method based on the depth probability generation model, a multi-level representation method is constructed, including a task embedding layer, a morphological latent variable generation layer and a voxel distribution generation layer, voxel robot morphology is generated, and the morphological pool and continuous natural selection sampling technology are optimized.
It realizes faster discovery of better-performing robot design patterns, improves sample utilization and multi-task design efficiency, and reduces dependence on traditional evolutionary algorithms.
Smart Images
Figure CN120010243A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of voxel robots, and in particular to a voxel robot structure optimization design method based on morphological statistical modeling. Background Art
[0002] Voxel robots are composed of a large number of voxel structural units, so compared with rigid robots with kinematic tree structures, they have higher structural deformation accuracy and complex morphology control capabilities, and are more flexible, stable and accurate in complex environments. In order to give full play to the advantages of voxel robots such as exquisite structure, flexible controllability, it is necessary to propose an effective morphology design method to obtain efficient and advantageous morphology design for tasks.
[0003] In order to solve this morphological design optimization problem, traditional robotics generally adopts the Type Synthesis method, that is, given some specific requirements and constraints, different types of mechanisms, such as joints and transmission mechanisms, are evaluated and compared to select the most suitable mechanism type for the task or working conditions. However, this method relies on expert experience and prior knowledge, and has limited adaptability to multi-task scenarios. The recent development of artificial intelligence has brought about changes in mechanical design research, and a research method has been derived that implements control optimization through reinforcement learning or evolutionary algorithms, obtains morphological evaluation, and then optimizes morphological design based on the evaluation results, namely the double-layer optimization design of "morphology-control collaborative optimization".
[0004] Existing technologies generally use evolutionary computing methods to search and optimize morphological space, such as Bayesian optimization and genetic algorithms. These methods are difficult to efficiently obtain the optimal solution when faced with the potentially large morphological search space of voxel robots. In addition, existing technologies generally only perform control optimization for specific tasks, and each morphological sample is only used once during the generational optimization process, resulting in low sample utilization and limited task generalization performance. 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 structure optimization design method based on morphological statistical modeling.
[0006] The technical solution of the present invention is as follows:
[0007] A voxel robot structure optimization design method based on morphological statistical modeling is provided, the method comprising:
[0008] Building and training a deep probabilistic generative model based on a multi-level representation approach, including a task embedding layer, a morphological latent variable generation layer, and a voxel distribution generation layer;
[0009] Using the generative model to perform morphological design on the target task includes: inputting a task code, generating a morphological latent variable, and generating a voxel robot morphology based on the morphological latent variable;
[0010] During the generation process, a morphological pool is used to store the morphological samples and their fitness of intergenerational sampling. The morphological samples are screened and optimized based on the continuous natural selection sampling technology, and the optimization results are fed back to the generation model.
[0011] Aiming at different task objectives, multi-task adaptive design is achieved by iteratively optimizing the generation model and sample selection strategy.
[0012] In one embodiment of the present invention, the continuous natural selection sampling technique is implemented by the following steps:
[0013] During the robot generation replacement process, each generation generates a morphology sample through the generation model, optimizes the controller according to the morphology sample and evaluates the fitness;
[0014] Put the morphological samples and the corresponding fitness scores into the morphological pool;
[0015] The morphological samples are sampled alternately from the morphological pool using probabilistic sampling, and stochastic gradient ascent is performed based on the extracted morphological samples to maximize the variational lower bound.
[0016] In one embodiment of the present invention, in the continuous natural selection sampling, the number of sampling and updating steps is adjusted according to the generation increase to control the convergence time within a preset range.
[0017] In one embodiment of the present invention, in the continuous natural selection sampling, the samples in the morphological pool are sorted according to the fitness score, the sampling probability is dynamically adjusted according to the fitness score, high fitness samples are preferentially selected, while ensuring that the probability of low fitness samples is non-zero, and the fitness score distribution is adjusted by controlling the selection pressure hyperparameter.
[0018] In one embodiment of the present invention, the deep probability generation model is constructed by a variational autoencoder module, comprising the following steps:
[0019] a) The task encoding embedding layer maps the task type to the continuous embedding space to generate the task embedding vector;
[0020] b) The multi-layer perceptron generates latent variable distribution parameters based on the task embedding vector, including the mean vector and the diagonal variance-covariance matrix;
[0021] c) In the generation process, latent variables are sampled from the latent variable distribution based on the multivariate normal distribution, and voxel distribution is generated through a multilayer perceptron;
[0022] d) During the optimization process, the variational lower bound is used to perform approximate posterior optimization on the generative model to improve the expressive power of the generative model.
[0023] In one embodiment of the present invention, it also includes: introducing exploration-exploitation rebalancing technology to generate a preset number of exploration samples in the generational sampling process to encourage the model to try diverse morphological designs, wherein the exploration samples are morphological samples whose similarity with the existing designs is below a preset threshold, and the utilization samples are morphological samples whose similarity with the existing designs is above a preset threshold.
[0024] In one embodiment of the present invention, the similarity between morphological samples is defined as follows:
[0025] Similarity(x,y)=(|x∩y|) / (|x∪y|)
[0026] Among them, Similarity(x,y) represents the similarity between sample x and sample y, and x and y are the voxel sets of two morphological samples respectively.
[0027] In one embodiment of the present invention, the exploration-exploitation rebalancing technique balances design efficiency and morphological diversity by adjusting the ratio of exploration samples to exploitation samples.
[0028] In one embodiment of the present invention, the generated morphological samples are suitable for multiple tasks, the task features are represented by task embedding vectors, and the morphological samples are represented by morphological embedding vectors. The task embedding vectors are generated by an embedding layer in a deep probabilistic generation model, and the morphological embedding vectors are obtained by a latent variable generation module. By inputting the task embedding vectors and the morphological embedding vectors into a shared multi-layer perceptron module, joint modeling and cross-task migration of task features and morphological features are achieved.
[0029] The main advantages of the technical solution of the present invention are as follows:
[0030] The voxel robot structure optimization design method based on morphological statistical modeling of the present invention constructs a deep probabilistic generation model, abstractly describes the relationship between tasks and morphological design, eliminates the dependence on traditional evolutionary algorithms, and makes full use of the multimodality, efficient calculation and pattern learning capabilities of the probabilistic generation model from different task advantage samples, which can more quickly discover robot design forms with better performance. The morphological pool is introduced to store samples and their fitness, and then sampling is performed based on "continuous natural selection" to maintain the wide applicability of the description, effectively improve sample utilization, and improve multi-task design efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] 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.
[0032] Figure 1 It is a flow chart of a voxel robot structure optimization design method based on morphological statistical modeling according to an embodiment of the present invention;
[0033] Figure 2 Schematic diagram of the construction and training process of the deep probability generation model in the voxel robot structure optimization design method based on morphological statistical modeling according to an embodiment of the present invention. DETAILED DESCRIPTION
[0034] 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.
[0035] The technical solution provided by the embodiments of the present invention is described in detail below with reference to the accompanying drawings.
[0036] The embodiment of the present invention provides a voxel robot structure optimization design method based on morphological statistical modeling, as shown in the attached Figure 1 As shown, including:
[0037] S1, build and train a deep probabilistic generative model. The generative model is based on a multi-level representation method including a task embedding layer, a morphological latent variable generation layer, and a voxel distribution generation layer boundary.
[0038] In an embodiment of the present invention, a probabilistic generative model based on deep learning is constructed to characterize the relationship between task type, morphological embedding and voxel distribution. The generative model maps task features to potential representations of morphological design through a hierarchical architecture, specifically including: taking the task type as input and generating the corresponding task representation. Based on the task representation, latent variables are generated to describe the potential distribution of morphological features. The structural representation of the voxel robot is generated using latent variables. During the model training process, the objective function is optimized by design to ensure that the generated morphological design can meet the requirements of different tasks. The generative model can choose a variety of implementation methods according to actual needs, such as variational autoencoders, generative adversarial networks, or other deep learning models suitable for characterizing complex distributions.
[0039] The use of generative models to characterize the relationship between tasks and morphology enables the structural optimization design method to adapt to a variety of task requirements. The model training process can capture the diversity of the morphological design space and the specific requirements of the task, improving design efficiency.
[0040] S2, using the generative model to perform morphological design for the target task, including: inputting task encoding, generating morphological latent variables, and generating voxel robot morphology based on the morphological latent variables.
[0041] After the construction and training of the generative model is completed, the morphological design of the target task is carried out through the following steps: First, the generative model is input according to the task type to obtain the potential representation of the morphological design. Then, a specific voxel robot morphology is generated based on the potential representation. The generated morphology can include key information such as the material type and distribution structure of the voxels. The generation process can use a variety of latent variable generation techniques, such as random sampling based on Gaussian distribution or other distribution generation methods suitable for the task scenario.
[0042] The generative model is used to complete the morphological design of the target task, automatically generate morphological design, avoid the traditional inefficient process of relying on manual experience, and can quickly adapt to multi-task requirements and generate highly adaptable voxel robot morphology.
[0043] S3, during the generation process, uses the morphological pool to store the morphological samples and their fitness of intergenerational sampling, and screens and optimizes the morphological samples based on the continuous natural selection sampling technology, and the optimization results are fed back to the generation model.
[0044] In the morphology design process, a morphology pool is introduced to store morphology samples and their fitness. The functions of the morphology pool include: storing morphology samples and their corresponding task fitness information, providing support for subsequent morphology optimization and evaluation. The diversity and adaptability of samples are ensured through dynamic updates of the morphology pool.
[0045] The introduction of the morphological pool avoids the inefficiency of single-use samples and greatly improves the optimization efficiency of morphological design; by storing task fitness information, it provides data support for subsequent multi-task optimization.
[0046] S4, for different task objectives, realizes multi-task adaptive design by iteratively optimizing the generation model and sample selection strategy.
[0047] By training the generative model and reasonably selecting morphological samples, this method can adapt to the design requirements of different task scenarios. Specifically: for different task inputs, the corresponding morphological design is generated through the generative model. According to the task objectives and morphological distribution characteristics, the parameters of the generative model are optimized to improve the efficiency of multi-task design.
[0048] In summary, the voxel robot structure optimization design method based on morphological statistical modeling provided by the embodiment of the present invention constructs a deep probabilistic generation model, abstractly describes the relationship between tasks and morphological design, eliminates the dependence on traditional evolutionary algorithms, and makes full use of the multimodality, efficient calculation and pattern learning capabilities of the probabilistic generation model from different task advantage samples, so as to more quickly discover robot design forms with better performance. A morphological pool is introduced to store samples and their fitness, and then sampling is performed based on "continuous natural selection" to maintain the wide applicability of the description, effectively improve sample utilization, and improve the efficiency of multi-task design.
[0049] The following is a detailed description of each step and the principles involved in the voxel robot structure optimization design method based on morphological statistical modeling provided by an embodiment of the present invention.
[0050] Considering that the voxel robot is formed by voxel units with multiple functions and performances arranged in a grid structure, the method provided in this embodiment of the present invention adopts multivariate normal distribution to model its dominant morphological distribution, and then uses a deep generative model for parameter inference and sampling. The model architecture can be divided into two parts: the probability generation process and the approximate posterior inference.
[0051] 1. Probability Generation Process
[0052] The generation process logically has a hierarchical structure, namely, task type, dominant morphology, and voxel type, which corresponds to the three-layer structure of task → morphology embedding → voxel matrix in the deep generation model, such as Figure 2 As shown in the dashed box on the left.
[0053] The input of the entire generation process is the encoding vector y of the task type, which follows a multinomial distribution and is converted into an embedding vector after passing through an embedding layer. Each task corresponds to a unique trainable embedding vector that captures the task characteristics. Input into two multi-layer perceptrons to generate mean vectors respectively and the diagonal variance-covariance matrix Where θ is the neural network parameter. Then from the multidimensional variable Gaussian distribution Finally, the latent variable is input into another multi-layer perceptron (denoted as ), and the voxel distribution x consisting of the material type of the voxel is calculated. Each voxel block of the voxel robot corresponds to a polynomial distribution. The entire generation process can be expressed by the following formula:
[0054]
[0055] The morphological samples generated based on the generative model are suitable for multiple tasks. When performing inter-task migration, the task features are represented by the task embedding vector, and the morphological samples are represented by the morphological embedding vector. The task embedding vector is generated by the embedding layer in the deep probabilistic generative model, and the morphological embedding vector is obtained by the morphological latent variable generation layer. By inputting the task embedding vector and the morphological embedding vector into a shared multi-layer perceptron module, the joint modeling and cross-task migration of task features and morphological features are achieved.
[0056] 2. Approximate Posterior
[0057] In order to construct a variational lower bound that is easy to optimize, we choose to use an approximate posterior to approximate the true posterior distribution of the latent variable instead of directly optimizing the original likelihood function. The obtained variational lower bound can be expressed as follows:
[0058]
[0059] Where γ represents the uniform distribution of task types, X Y represents the distribution of favorable forms corresponding to task y, p θ Represents the generation process mentioned above, represents the approximate posterior.
[0060] 3. Continuous Natural Selection Sampling
[0061] Next, under the two-level optimization framework of morphology design and control, a deep generative model training method of "continuous natural selection" using morphology pool and generational replacement optimization is adopted. Specifically, during the robot generational replacement process, each generation generates a robot population (morphology set) through a generative model, optimizes its controller and evaluates its fitness. Then, we put these designs and fitness scores into the morphology pool. Subsequently, a probabilistic sampling method based on the fitness score probability formula is designed to sample robots from the morphology pool in turn, and perform stochastic gradient ascent based on the extracted robots to maximize the variational lower bound. In each generation, the sampling and update steps are performed multiple times, and the number of steps increases linearly with the number of generations to prevent premature convergence. The fitness-based probability sampling formula is defined as follows:
[0062]
[0063] in represents the i-th robot corresponding to task h, express , H is the total number of tasks, K is the total number of robots in the morphology pool corresponding to task h, and τ is a hyperparameter that controls the intensity of the selection pressure. When τ is high, the method prefers the best performing robots, while when τ is low, the method pays attention to some robots with slightly worse performance.
[0064] In this way, robots with better performance will be more likely to be selected, thus ensuring that the generated robots are optimized in the direction of gradual improvement. At the same time, because this method does not directly set a threshold to judge the quality of the robot, but continuously adjusts the probability of each robot design appearing in the sample according to its fitness, it can avoid missing favorable designs or mixing unfavorable designs, and enable us to make better use of all the robot designs we have evaluated. Therefore, this process can be regarded as a continuous and flexible natural selection method.
[0065] 4. Explore and utilize rebalancing technology
[0066] Since the generative model repeatedly samples and fits robot samples from the morphology pool, there may be problems such as the newly generated morphology being too conservative and similar individuals recurring during the generational replacement process, which is not conducive to maintaining population diversity and design robustness. In order to overcome this problem, this method proposes a technique called exploration-exploitation rebalancing. Specifically, assuming that a morphology has no less than s% of the same voxels as the existing design, the morphology is called an "exploitation sample", and vice versa, it is called an "exploration" sample. The deep generative model is then repeatedly applied in each generation until a predetermined number of "exploration samples" are obtained, thereby encouraging the model to try more diverse robot morphology designs and preventing it from relying too much on favorable robot designs in the morphology pool, such as Figure 2 As shown in the middle part. It can be understood that the higher the proportion of exploration samples, the more the generative model is encouraged to explore morphological diversity. The lower the proportion of exploration samples, the higher the design efficiency of the generative model. In actual use, the ratio of exploration samples to utilization samples can be adjusted to balance design efficiency and morphological diversity.
[0067] The value of s can be designed by those skilled in the art according to actual working conditions, for example, it can be 10, 20, etc., and is not specifically limited in the embodiments of the present invention. The similarity between morphological samples is defined as follows:
[0068] Similarity(x,y)=(|x∩y|) / (|x∪y|)
[0069] Among them, Similarity(x,y) represents the similarity between sample x and sample y, and x and y are the voxel sets of two morphological samples respectively.
[0070] 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.
[0071] 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 structure optimization design method based on morphological statistical modeling, characterized in that: include: Building and training a deep probabilistic generative model based on a multi-level representation approach, including a task embedding layer, a morphological latent variable generation layer, and a voxel distribution generation layer; Using the generative model to perform morphological design on the target task includes: inputting a task code, generating a morphological latent variable, and generating a voxel robot morphology based on the morphological latent variable; During the generation process, a morphological pool is used to store the morphological samples and their fitness of intergenerational sampling. The morphological samples are screened and optimized based on the continuous natural selection sampling technology, and the optimization results are fed back to the generation model. Aiming at different task objectives, multi-task adaptive design is achieved by iteratively optimizing the generation model and sample selection strategy.
2. The voxel robot structure optimization design method based on morphological statistical modeling according to claim 1 is characterized in that: The continuous natural selection sampling technique is implemented by the following steps: During the robot generation replacement process, each generation generates a morphology sample through the generation model, optimizes the controller according to the morphology sample and evaluates the fitness; Put the morphological samples and the corresponding fitness scores into the morphological pool; The morphological samples are sampled alternately from the morphological pool using probabilistic sampling, and stochastic gradient ascent is performed based on the extracted morphological samples to maximize the variational lower bound.
3. The voxel robot structure optimization design method based on morphological statistical modeling according to claim 2 is characterized in that: In the continuous natural selection sampling, the number of sampling and updating steps is adjusted according to the generation increase to control the convergence time within a preset range.
4. The voxel robot structure optimization design method based on morphological statistical modeling according to claim 2 is characterized in that: In the continuous natural selection sampling, the samples in the morphological pool are sorted according to the fitness score, and the sampling probability is dynamically adjusted according to the fitness score, and high fitness samples are preferentially selected while ensuring that the probability of low fitness samples is non-zero. The fitness score distribution is adjusted by controlling the selection pressure hyperparameter.
5. The voxel robot structure optimization design method based on morphological statistical modeling according to claim 1 is characterized in that: The deep probability generation model is constructed through a variational autoencoder module, which includes the following steps: a) The task encoding embedding layer maps the task type to the continuous embedding space to generate the task embedding vector; b) The multi-layer perceptron generates latent variable distribution parameters based on the task embedding vector, including the mean vector and the diagonal variance-covariance matrix; c) In the generation process, latent variables are sampled from the latent variable distribution based on the multivariate normal distribution, and voxel distribution is generated through a multilayer perceptron; d) During the optimization process, the variational lower bound is used to perform approximate posterior optimization on the generative model to improve the expressive power of the generative model.
6. The voxel robot structure optimization design method based on morphological statistical modeling according to any one of claims 1 to 5, characterized in that: It also includes: introducing exploration-utilization rebalancing technology to generate a preset number of exploration samples in the generational sampling process to encourage the model to try diverse morphological designs, where the exploration samples are morphological samples whose similarity with the existing designs is below a preset threshold, and the utilization samples are morphological samples whose similarity with the existing designs is above a preset threshold.
7. The voxel robot structure optimization design method based on morphological statistical modeling according to claim 6 is characterized in that: The similarity between morphological samples is defined as follows: Similarity(x,y)=(|x∩y|) / (|x∪y|) Among them, Similarity(x,y) represents the similarity between sample x and sample y, and x and y are the voxel sets of two morphological samples respectively.
8. The voxel robot structure optimization design method based on morphological statistical modeling according to claim 7 is characterized in that: The exploration-exploitation rebalancing technique balances design efficiency and morphological diversity by adjusting the ratio of exploration samples to exploitation samples.
9. The voxel robot structure optimization design method based on morphological statistical modeling according to claim 1 is characterized in that: The generated morphological samples are suitable for multiple tasks, the task features are represented by task embedding vectors, and the morphological samples are represented by morphological embedding vectors. The task embedding vectors are generated by the embedding layer in the deep probabilistic generation model, and the morphological embedding vectors are obtained by the latent variable generation module. By inputting the task embedding vectors and the morphological embedding vectors into a shared multi-layer perceptron module, the joint modeling and cross-task migration of task features and morphological features are realized.