Dexterous hand structure parameter and control strategy collaborative optimization method and system

Abstract

This invention discloses a method and system for co-optimizing structural parameters and control strategies of a dexterous hand. The method includes: constructing a baseline model of the dexterous hand and defining its parameters; automatically updating the structural description file and geometric model file based on a pre-constructed parameter-model field mapping relationship, and simultaneously deriving derived dynamic parameters such as the centroid position and inertia tensor to generate an updated simulation model; constructing an execution network including a general state network, a structurally differentiated network, and an action strategy network, and combining a value network with structural consistency regularization constraints and a structure-aware generalized advantage estimation, training the model based on a proximal strategy optimization algorithm until the task success rate under the current structure converges; training a Gaussian process surrogate model based on the convergence strategy and iteratively searching for the optimal structural parameters through a collection function, comparing the performance of candidate structures to update the baseline model, and iteratively iterating the inner and outer loop optimization process, outputting the optimal parameters and strategy after reaching the target.

Description

Technical Field

[0001] This invention belongs to the field of robot motion control technology, and more specifically, relates to a method and system for the collaborative optimization of dexterous hand structural parameters and control strategies. Background Technology

[0002] As a key actuator in service robots, bionic robots, and industrial automation equipment, dexterous hands are highly challenging to design and control due to their multiple degrees of freedom, multiple joints, and complex drive structures.

[0003] The design and optimization of existing dexterous hands typically adopts a phased approach of "structure first, control later." In terms of structural design, it mainly relies on engineers' experience-based parameter trial and error and finite element verification, lacking systematic and adaptive design methods. This results in long design cycles, high trial-and-error costs, and difficulty in achieving adaptive optimization of structural parameters. Regarding control strategies, deep reinforcement learning has been gradually applied to training dexterous hands for complex maneuvers in recent years. However, its high dependence on high-fidelity simulation environments leads to high training costs, and the trained strategies exhibit significant simulation-reality gaps when transferred to real systems, resulting in high transfer costs, poor stability, and difficulty in direct application.

[0004] To address these issues, some solutions attempt to introduce digital twin technology to construct a virtual mapping model of the dexterous hand, aiming to bridge the gap between virtual training and physical execution. However, most existing digital twin systems only achieve a static mirror image of the physical entity and do not form a closed-loop feedback mechanism with structural optimization algorithms and reinforcement learning training frameworks. Once the structural parameters change, the model still needs to be rebuilt and the control strategy training needs to be restarted, making it impossible to achieve joint optimization of the structural space and control space.

[0005] Furthermore, in the few attempts at structure-control co-optimization, a naive search approach of "training a separate set of policies for each set of structures" is often adopted. This not only consumes huge computational resources, but also easily leads to policy gradient conflicts due to the differences in the optimal action distributions corresponding to different structures, making training difficult to converge and severely restricting the adaptive deployment capability of dexterous hands in complex task scenarios.

[0006] The aforementioned problems also lead to the fact that structural design and control strategies are often disconnected in the current development of dexterous hands. This requires repeated strategy training and structural adjustments from scratch, resulting in low development efficiency, high resource consumption, and difficulty in obtaining globally optimal design and control solutions.

[0007] Therefore, there is an urgent need for a method that can achieve integrated and collaborative optimization of dexterous hand structural parameters and control strategies within a single framework, in order to improve the level of design automation, shorten the R&D cycle, and ensure optimal performance. Summary of the Invention

[0008] To address the aforementioned problems in existing technologies, this invention provides a method and system for the collaborative optimization of dexterous hand structural parameters and control strategies. It employs an embedded Bayesian-deep reinforcement learning closed-loop framework, deeply integrating outer-loop Bayesian optimization with inner-loop policy gradient optimization. By constructing a separate execution network of a general state network and a structurally differentiated network, it decouples cross-structure general control laws from specific structural compensation characteristics. Combined with structure-aware generalized advantage estimation and value network structural consistency regularization constraints, it significantly reduces sample requirements and computational consumption while ensuring training stability. Simultaneously, it includes an open automatic parameter mapping mechanism, enabling automatic synchronous updates of optimization algorithm output parameters to structural description files and geometric model files, as well as automatic derivation of derived dynamic parameters. Furthermore, it utilizes detachable joint modules to achieve low-cost and rapid mapping of simulation optimization results to the physical system. Thus, within a single framework, it achieves integrated collaborative evolution of dexterous hand morphological parameters and control strategies, effectively solving the problems of low optimization efficiency, severe gradient conflicts, and poor physical consistency in traditional methods. This enables high-performance, adaptive automated design and efficient iteration of dexterous hands.

[0009] To achieve the above objectives, one aspect of the present invention provides a method for co-optimizing the structural parameters and control strategy of a dexterous hand, comprising the following steps: S1: Construct a baseline model of the dexterous hand and define its parameters; model each finger joint of the dexterous hand using a symmetrical geometric structure, and constrain the optimizable parameters into three core geometric parameters: joint length, width, and spatial position of the joint relative to the palmar plane, forming an open structural parameter vector; each joint is designed as a detachable modular structure. S2: Based on the pre-built parameter-model field mapping relationship, the corresponding nodes in the dexterous hand structure description file are traversed and updated respectively, and the geometric model file is transformed into a parameterized intermediate representation and replaced with parameters. The centroid position and inertial tensor derived dynamic parameters derived from geometric changes are updated synchronously to generate the updated dexterous hand simulation model. S3: Import the updated dexterous hand simulation model into the training environment, construct an execution network including a general state network, a structural alienation network and an action policy network, combine a value network with structural consistency regularization constraints and structural awareness generalized advantage estimation, and train based on the proximal policy optimization algorithm until the task success rate under the current structure converges to a preset threshold. S4: Randomly sample multiple sets of candidate structure parameters and evaluate the corresponding task success rates. Construct a sampling dataset containing structure parameters and success rates. Train a Gaussian process surrogate model and select the next optimal candidate structure parameter through a sampling function. Input the candidate parameters into the simulation environment and use the converged strategy to obtain the true success rate. S5: If the actual success rate is better than the historical best value, then the new parameter is used as the baseline structural parameter, and the process returns to step S2 to update the dexterous hand model, and steps S3 to S4 are repeated; if it is not better, then the parameter and the corresponding success rate are added to the sampling dataset, the Gaussian process surrogate model is updated and candidate parameters are reselected, and the process is iterated until the global success rate target is reached. The optimal structural parameter and the corresponding optimal control strategy are output, and the optimal structural parameter is mapped to the detachable joint module to realize the rapid conversion of simulation optimization results into a physical dexterous hand.

[0010] Furthermore, the dexterous hand described in step S1 includes the thumb, index finger, and middle finger, and adopts a hybrid drive method combining a linear actuator and a rotary actuator; The fingers of the dexterous hand are connected to the palm base through joints and are driven by multiple actuators to achieve multiple degrees of freedom. Each finger of the dexterous hand is equipped with an actuator embedded in the palm; the thumb achieves flexion movement at the proximal and distal ends of the palm through two independent actuators, the proximal ends of the index and middle fingers achieve flexion movement through linear actuators and transmission mechanisms, while the middle and distal ends of the palm achieve flexion control through independent rotary actuators.

[0011] Furthermore, the open structural parameter vector in step S1 has a dimension of 23, including the length and width of the proximal, middle, and distal palmar joints of the index and middle fingers, the length and width of the proximal and distal palmar joints of the thumb, and the spatial position parameters of the preset joints or joints of the index, middle, and thumb relative to the palmar plane.

[0012] Further, step S2 includes: The structural parameters in step S1 are preprocessed to establish a parameter-model field mapping table; the parameter-model field mapping table includes a structural description file field mapping table and a geometric model file field mapping table. After the optimization algorithm outputs the structural parameter vector to be updated, the system parses the original structural description file, constructs a node tree, and traverses and replaces the node fields corresponding to the target parameters according to the mapping table, only modifying the parts directly related to the optimization parameters, keeping the model topology and assembly relationship unchanged; For geometric model files, first convert them into an editable intermediate representation, then iterate through and update the relevant geometric feature dimensions according to the geometric parameter mapping table, and finally restore the modified intermediate representation file to a standard geometric model file. After the model is updated, the derived dynamic parameters, including joint mass, center of mass position, and inertia tensor, are automatically calculated and updated based on the changes in joint geometry.

[0013] Furthermore, in step S3, the general state network receives the environment and ontology state inputs, the structural alienation network receives the current structural parameter vector input, and the action strategy network fuses the outputs of both to generate an action probability distribution.

[0014] Furthermore, the total loss function of the value network described in step S3 is expressed as follows:

[0015] in, This represents the time steps sampled during training. Perform an empirical average; That is, the expected value output of the value network; It is the value in the actual execution process; For structural consistency regularization terms Weighting coefficients; It is a structural consistency regularization term; Furthermore, structural consistency regularization terms The expression is:

[0016] in, Represents the state in the training samples and structural parameter combination index Empirical averaging is performed to statistically analyze the differences in the output of value functions corresponding to similar structural parameters under different states; This represents the proximity threshold of structural parameters, used to determine the proximity of two sets of structural parameters. and Do they belong to similar structures? To measure structure and Weights of similarity ; This is a hyperparameter used to control the rate of weight decay.

[0017] Furthermore, the expression for the structure-aware generalized advantage estimation described in step S3 is:

[0018] in, For the first Generalized dominance estimates under structural parameters; It is the offset of the time step; subscript + Relative to the current time The future At that moment; It is a discount factor, and its value is usually between [0,1]. For GAE attenuation parameters; For the current moment The future The timing difference error at each moment; For structural weighting factors; The expression for the structure weight factor is:

[0019] in, For the first Timing difference error under various structural parameters , It is a moment The reward value; It is an estimate of the value of the current state; This is a value estimate for the next state.

[0020] Furthermore, the loss function expression for the network executed in step S3 is as follows:

[0021] in, The target loss function of the PPO-Clip algorithm; This represents the time steps sampled during training. Perform an empirical average; For strategy ratio; It is the dominance function.

[0022] Furthermore, the Gaussian process model expression in step S4 is: ,in, The function to be optimized is a black-box function. These are structural parameters, containing all possible values ​​for all parameters; It is a mean function; For kernel functions; The acquisition function used in step S4 is the expected boost function.

[0023] A second aspect of the present invention provides a system for the coordinated optimization of dexterous hand structural parameters and control strategies, used to implement the aforementioned method for the coordinated optimization of dexterous hand structural parameters and control strategies, comprising: The benchmark model construction and parameterization module is used to construct a benchmark model of the dexterous hand and define its parameters. The symmetrical geometric structure model is used for each finger joint of the dexterous hand. The optimizable parameters are constrained into three types of core geometric parameters: joint length, width and spatial position of the joint relative to the palm plane, forming an open structural parameter vector. Each joint is designed as a detachable modular structure. The open parameter automatic mapping module is used to establish an open parameter automatic mapping mechanism. Based on the pre-built parameter-model field mapping relationship, it traverses and updates the corresponding nodes in the dexterous hand structure description file, and performs parameterized intermediate representation conversion and parameter replacement on the geometric model file. It also updates the centroid position and inertial tensor derived dynamic parameters derived from geometric changes, and generates the updated dexterous hand simulation model. An embedded collaborative optimization closed-loop network module is used to perform reinforcement learning training of the inner-loop control strategy. The updated dexterous hand simulation model is imported into the training environment to construct an execution network including a general state network, a structural alienation network, and an action policy network. The general state network receives environmental and ontological state inputs, the structural alienation network receives the current structural parameter vector input, and the action policy network fuses the outputs of both to generate an action probability distribution. Combined with a value network with structural consistency regularization constraints and structural awareness generalized advantage estimation, the system is trained based on a proximal policy optimization algorithm until the task success rate under the current structure converges to a preset threshold. The Bayesian outer loop optimization module performs Bayesian optimization of the outer loop structure parameters, randomly samples multiple sets of candidate structure parameters and evaluates the corresponding task success rates, constructs a sampling dataset containing structure parameters and success rates, trains a Gaussian process surrogate model, selects the next optimal candidate structure parameters through a sampling function, calls an open parameter automatic mapping mechanism to generate the corresponding dexterous hand simulation model, inputs the candidate parameters into the simulation environment, and obtains the true success rate using the converged control strategy. The iterative control and termination judgment module is used to coordinate the interaction logic between the inner and outer loops and control the start and stop of the optimization process. If the actual success rate is better than the historical best value, the new parameters are used as the benchmark structural parameters, and the process returns to step S2 to update the dexterous hand model. Steps S3 to S4 are repeated. If the success rate is not better than the historical best value, the parameters and the corresponding success rate are added to the sampling dataset, the Gaussian process surrogate model is updated, and candidate parameters are reselected. The process is iteratively run until the global success rate target is reached, and the optimal structural parameters and the corresponding optimal control strategy are output. The optimal structural parameters can be mapped to the detachable joint module to realize the rapid conversion of simulation optimization results into a physical dexterous hand.

[0024] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) This invention integrates Bayesian optimization (outer loop) and deep reinforcement learning (inner loop) by constructing an embedded structure-control collaborative optimization closed-loop network framework. This framework breaks the traditional design pattern of "determining the structure first and then adjusting the control". By exploring the optimal structural parameters globally in the outer loop and quickly adapting the optimal control strategy in the inner loop, it realizes the integrated and adaptive joint optimization of the dexterous hand's structural parameters and control strategy. Through the collaborative operation of the global exploration of structural parameters in the outer loop and the policy gradient optimization in the inner loop, it effectively solves the problem of the separation between structural design and control strategy in traditional methods, which requires repeated training from scratch.

[0025] (2) This invention innovatively constructs a separate execution network architecture consisting of a general state network and a structurally differentiated network, and uses both as pre-networks of the action policy network. The general policy network takes only state parameters as input to learn the basic control laws that are invariant across structures. The structurally differentiated network takes only structural parameters as input to characterize the control compensation characteristics under different structural conditions. Through this decomposition method, the complex control policy learning problem that originally varied with both state and structural parameters is decomposed into two sub-problems: "cross-structure shared control learning" and "structure-related control compensation". This avoids repeated training of the general control policy under different structural parameters and significantly reduces the optimization complexity and sample requirements of the inner-loop reinforcement learning network. The above network decomposition mechanism also indirectly alleviates the gradient conflict problem caused by the difference in optimal actions due to different structural parameters at the optimization level, making the policy gradient more stable during parameter updates, which is conducive to the rapid convergence and generalization ability of the execution network. Compared with the traditional method of training control policies independently for each set of structural parameters, this invention significantly improves training efficiency and reduces computational consumption.

[0026] (3) This invention proposes a structure-aware generalized advantage estimation and a value network structure consistency regularization constraint mechanism. The structure-aware generalized advantage estimation introduces an adaptive weight factor based on the variance of the value estimation, making the policy update more conservative when the structural uncertainty is strong and more aggressive when the policy is stable. The regularization constraint forces the value estimation of similar structures to remain continuous, avoiding non-physical jumps. The two work together to significantly improve the convergence stability and generalization ability of reinforcement learning training during the process of structural parameter changes.

[0027] (4) The open parameter automatic mapping mechanism proposed in this invention realizes the automatic and accurate synchronous update of the output parameters of the optimization algorithm to the structural description file and the geometric model file by establishing a parameter-model field mapping table and a geometric file standardization conversion process. At the same time, it automatically derives and updates derived dynamic parameters such as the centroid and inertia tensor based on geometric changes, ensuring the consistency of the model in topology and physical properties, and completely eliminating the tedious process of manual modeling and adjustment.

[0028] (5) By designing a detachable joint module and a hybrid drive benchmark model, the present invention constrains the optimizable parameters into three core variables: length, width and position, which greatly compresses the parameter space. The optimized parameters can be directly mapped to the physical module for replacement without reprocessing the entire hand, which significantly reduces the manufacturing and debugging costs of multi-scheme physical verification and makes the rapid customization design of dexterous hands possible.

[0029] (6) This invention proposes a structural parameter update mechanism based on a converged strategy. The outer-loop Bayesian optimization directly uses the converged control strategy as the evaluation basis, uses a Gaussian process surrogate model to fit the parameter-performance mapping, and intelligently balances exploration and utilization through the acquisition function. This avoids the large number of invalid evaluations required by blind grid search or genetic algorithm, and can efficiently lock the global optimal structural parameters in a smaller number of iterations, ensuring that the dexterous hand design scheme output in the end has the best comprehensive operation performance. At the same time, when sampling and evaluating the performance of different combinations of structural parameters, this mechanism does not require retraining from scratch, so that each update of the structural parameters is based on the existing strategy model, which significantly reduces the computational cost and time overhead of the outer-loop optimization, and ensures the continuity and stability of the strategy in the process of structural parameter changes. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating a method for the coordinated optimization of structural parameters and control strategies of a dexterous hand according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the algorithm flow of the optimization strategy in a method for co-optimizing structural parameters and control strategies of a dexterous hand according to an embodiment of the present invention; Figure 3 A baseline three-finger dexterous hand outline constructed for embodiments of the present invention; Figure 4 This is a schematic diagram of a dexterous hand structure parameter and control strategy collaborative optimization system according to an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0032] like Figure 1 and Figure 2 As shown, one aspect of the present invention provides a method for co-optimizing the structural parameters and control strategy of a dexterous hand, comprising the following steps: S1. Construct a baseline model of the dexterous hand and define its parameters. Model each finger joint of the dexterous hand using a symmetrical geometric structure. Constrain the optimizable parameters into three core geometric parameters: joint length, joint width, and the spatial position of the joint relative to the palmar plane. Form an open structural parameter vector and design each joint as a detachable modular structure. S2. Establish an open parameter automatic mapping mechanism. In response to the structural parameter vector to be updated output by the optimization algorithm, based on the pre-built parameter-model field mapping relationship, traverse and update the corresponding nodes in the dexterous hand structure description file, and perform parameterized intermediate representation conversion and parameter replacement on the geometric model file. Simultaneously update the centroid position and inertial tensor derived dynamic parameters derived from geometric changes to generate the updated dexterous hand simulation model. S3. Execute the inner loop control strategy reinforcement learning training, import the updated dexterous hand simulation model into the training environment, and construct an execution network including a general state network, a structural alienation network and an action policy network. The general state network receives the environment and ontology state input, the structural alienation network receives the current structure parameter vector input, and the action policy network merges the outputs of the two to generate an action probability distribution. Combine the value network with structural consistency regularization constraints and the structure-aware generalized advantage estimation, and train based on the proximal policy optimization algorithm until the task success rate under the current structure converges to a preset threshold. S4. Perform Bayesian optimization of the outer loop structure parameters. Based on the convergent control strategy, randomly sample multiple sets of candidate structure parameters and evaluate the corresponding task success rates. Construct a sampling dataset containing structure parameters and success rates, train a Gaussian process surrogate model, and select the next optimal candidate structure parameters through the sampling function. Input the candidate parameters into the simulation environment and use the converged strategy to obtain the true success rate. S5. Compare the performance of candidate structures, select the best to update the baseline model, and iterate the inner and outer loop optimization process. After reaching the target, output the optimal parameters and strategy, and implement the physical modular dexterous hand. Specifically, if the actual success rate is better than the historical best value, the new parameters are used as the baseline structural parameters, return to step S2 to update the dexterous hand model, and repeat steps S3 to S4. If it is not better, add the parameter and the corresponding success rate to the sampling dataset, update the Gaussian process surrogate model, and reselect candidate parameters. Iterate until the global success rate target is reached, output the optimal structural parameters and the corresponding optimal control strategy, and map the optimal structural parameters to the detachable joint module to realize the rapid conversion of simulation optimization results into a physical dexterous hand.

[0033] This invention achieves the co-evolution of dexterous hand form and control by embedding Bayesian optimization and deep reinforcement learning. It solves the problems of the separation between structural design and control strategy, the need for repeated training from scratch, and the high cost of simulation-to-real-world transfer in traditional design. It significantly reduces computing power consumption and physical verification costs, and improves training stability and optimization efficiency.

[0034] The steps of the method of the present invention will be described in detail below.

[0035] (1) Construct a dexterous hand baseline model and define its parameters. A baseline model of a dexterous hand for structure-control co-optimization is constructed. The dexterous hand includes a thumb, index finger, and middle finger, each finger connected to a palm base via joints, and driven by multiple actuators to achieve multi-degree-of-freedom actuation, thus meeting the demands of diverse operational tasks for dexterous hand mobility and control independence. It should be noted that the dexterous hand of this invention is not limited to three fingers. In a preferred embodiment of this invention, refer to... Figure 3 Each finger is connected to the palm base and is driven by 11 actuators to achieve 11 degrees of freedom; In terms of the drive structure, this invention adopts a hybrid drive method combining linear actuators and rotary actuators, enabling each finger to have a degree of freedom of movement in the direction perpendicular to the palm plane, while also achieving independent flexion movements of multiple joints; improving the dexterity and operational ability of the hand while ensuring structural compactness; specifically, each of the three fingers is equipped with an actuator embedded in the palm, enabling rotational freedom in the direction perpendicular to the palm plane; the thumb achieves flexion movements at the proximal and distal ends of the hand through two independent actuators, the proximal ends of the index and middle fingers achieve flexion movements through linear actuators and transmission mechanisms, while the middle and distal ends of the hand achieve flexion control through independent rotary actuators, thus enabling the dexterity to have high mobility and control independence while ensuring structural compactness; In terms of structural parameter design, to reduce the parameter dimension of the structure-control co-optimization problem and improve the convergence of the optimization process, this invention adopts a symmetrical geometric structure model for each joint of each finger in the dexterous hand, abstracting the originally complex geometric shape into a regular cuboid or cylinder. This constrains the optimizable parameters of each joint, ensuring that only two core geometric quantities, length and width, are retained for the optimizable structural parameters of each joint. This significantly compresses the parameter space at the structural level and eliminates redundant design variables. By effectively reducing the parameter dimension of the structure-control co-optimization problem in the dexterous hand, the stability and convergence efficiency of the optimization process are improved. At the same time, this invention designs each joint of each finger as a detachable modular structure. When the optimization algorithm outputs a new combination of structural parameters, only the joint modules related to the parameter changes need to be replaced to complete the overall structural update of the dexterous hand, without the need to reprocess or reconstruct the entire hand, greatly reducing the iteration cost of the physical prototype. Based on the aforementioned dexterous hand baseline model, to support subsequent structure-control co-optimization tasks, this invention parametrically models the variable structural elements of the dexterous hand and defines them as an open structural parameter vector: ,in, For length is The structural parameter vectors are configured such that each parameter corresponds to the geometric dimensions of different finger joints and their spatial position relative to the palmar plane. By uniformly numbering and classifying each parameter, the structural parameters can be directly called by the optimization algorithm in vector form, thereby achieving efficient connection between structural design variables and the control strategy optimization process. Based on the aforementioned dexterous hand baseline model and corresponding parametric modeling, two types of standardized files are exported to support automated processing: structural description files and geometric model files. Structural description file: used to describe each bone segment, joint, and their connection relationships (e.g., URDF or MJCF format); by parsing the parameter definition rules in the structural description file, a set of variable parameters is extracted. This clarifies the position of the parameters in the model logic; Geometric model files: used to describe the three-dimensional shape of each component (such as STEP or STL format); to facilitate automated code modification, this invention first converts the geometric model files into parametric geometric intermediate representation files (such as parametric models defined based on OpenSCAD or Python scripts), and then forms a set of geometrically variable parameters through corresponding parameter definition rules: , where each parameter All of them are standardized into data units containing "structural element names" and "parameter values" (such as joint length, finger curvature, material stiffness, etc.), which together constitute a well-defined "open parameter set"; In a preferred embodiment of the present invention This includes the length and width of the proximal, middle, and distal palmar phalanges of the index and middle fingers, the length and width of the proximal and distal palmar phalanges of the thumb, and the spatial position parameters of the pre-defined phalanges or joints of the index, middle, and thumb relative to the palmar plane; their specific definitions are shown in Table 1: Table 1 Definition of Optimizable Structural Parameters

[0036] (2) Establish an open parameter automatic mapping mechanism Building upon the aforementioned dexterous hand baseline model, this invention further proposes an open-ended automatic parameter mapping mechanism for structural optimization. This mechanism enables automatic matching and updating between the structural parameter vectors output by the optimization algorithm and the dexterous hand model file. Using structural description files and geometric representation files as core carriers, this mechanism achieves synchronous modification and consistent evolution of structural parameters across different model representations through a unified parameter field mapping and file parsing update process.

[0037] Step S2 aims to establish an automated interface to accurately and losslessly map the abstract parameter vectors output by the optimization algorithm back to the physical model file of the dexterous hand, achieving synchronous evolution of "parameters-model"; specifically, it includes the following sub-steps: S21: Parameter Preprocessing and Construction of Parameter-Model Field Mapping Table For the set of structural parameters output by the optimization algorithm, preprocessing is first performed to establish a parameter-model field mapping table, clarifying the one-to-one correspondence between each component in the parameter vector and the corresponding field in the structural description file (as shown in Table 2). The parameter-model field mapping table includes a structural description file field mapping table and a geometric model file field mapping table. Using the parameter-model field mapping table, the system updates the structural parameter vector to be updated. The numerical values ​​in the data are matched with specific nodes (such as joint positions and centroid coordinates of bone segments) in the structural description file, and the data structure is adjusted to form a standardized parameter mapping table. Each element Each includes the name of the corresponding structural element in the structural description file and the corresponding optimized parameter values, thus providing a unified interface for subsequent automatic file-level modification. Table 2. Parameter-Structure Description Field Mapping Relationship Table

[0038] S22: Parsing and Updating Structure Description Files When the optimization algorithm of the embedded collaborative optimization closed-loop network outputs a set of structural parameter vectors to be updated Then, the system reads the current dexterous hand's structural description file, parses it, and constructs a node tree structure; then, it traverses the node tree according to the mapping table and, according to the parameter mapping table... Locate the target node; the update process adopts the principle of minimal modification: only the fields listed in the mapping table that are directly related to the optimization parameters are replaced with values, while keeping the other topological structure, connection relationship and comments of the file unchanged, so as to ensure the accuracy of the structural parameter update and avoid damaging the overall model topology and assembly relationship; S23: Normalization and Reconstruction of Geometric Model Files To address the problem that geometric model files (such as STEP / STL) are difficult to modify directly, this invention further proposes a method for geometric file normalization conversion and reverse reconstruction: Normalization conversion: Convert the original geometric model file into a parametric intermediate geometric representation file with explicit parameter definition capabilities, so that the size, curvature and relative position of each geometric element can be described in parametric form, and thus can be directly modified through code; Parameter traversal and update: by establishing geometric parameters With structural parameters Table of Relationships and Changes between Them (As shown in Table 3), traverse the source code of the parametric geometry intermediate representation file, update all corresponding parameters, and generate a new parametric geometry intermediate representation file; Reverse reconstruction: After the modification is completed, the updated parametric geometric intermediate representation file is recompiled / exported as a standard 3D geometric model file, thereby realizing the automatic reconstruction of the geometric model in the parameter space; Table 3. Parameter-Geometric Feature Correlation Changes

[0039] S24: Automatic Derivation of Derived Dynamic Parameters After the structural description file and geometric representation file are updated synchronously, the dexterous hand model is refreshed as a whole using the updated model file, while keeping the pre-set dynamics and actuator parameters unchanged. At the same time, based on the changes in structural parameters, the derived dynamic parameters caused by the geometric changes are automatically updated through dynamic derivation relationships, including but not limited to key parameters such as the center of mass position and inertia tensor, thereby ensuring the consistency and physical rationality of the updated model at the structural and dynamic levels. In a specific embodiment of the present invention, the system automatically recalculates the mass, center of mass position, and inertia tensor of each joint through volume integration based on the new geometric dimensions of the joints. The formulas for calculating the mass of each joint are as follows:

[0040] in, Total mass of the joints (unit: kg); For integral infinitesimal elements Spatial coordinates in the current coordinate system; Density function of bone and joint material in space (unit: kg / m³) 3 The total weight of a bone segment is obtained by calculating the sum of all the tiny mass elements within its volume. Center of mass location: the coordinates of the joint's center of mass in a three-dimensional coordinate system. (Unit: m) The expression is: , ,

[0041] Inertia tensor It is a 3×3 matrix used to describe an object's ability to resist rotation about different axes (moment of inertia): , The formula for calculating the diagonal elements (principal moments of inertia) in a matrix: , The formula for calculating the off-diagonal elements (product of inertia) in a matrix (reflecting the asymmetry of mass distribution): ; Through the above-mentioned open parameter automatic mapping and model file collaborative update mechanism, this invention realizes the end-to-end automatic conversion between the output parameters of the optimization algorithm and the dexterous hand structure model, so that the update process of the structure parameters is completely freed from the manual modeling and manual adjustment links, which significantly improves the efficiency, stability and consistency of the structure optimization iteration, and provides key technical support for the engineering implementation and large-scale application of dexterous hand structure-control closed-loop optimization. (3) Reinforce learning training by executing the inner loop control strategy through the embedded collaborative optimization closed loop system. Step S3 aims to fix the structural parameters. Under these conditions, train a high-performance control strategy. ; Step S3 specifically includes: S31: Construct a separate execution network The execution network includes a general state network, a structurally differentiated network, and an action policy network; The execution network is divided into new execution networks and old execution networks. The new execution network is the execution network after the parameter update after one iteration. In the next iteration, it will be used as the old execution network to obtain an action probability distribution for calculating the loss function. Define the set of network parameters in the execution network as : The general state network (in a specific embodiment of the present invention, the general state network is a multi-layer fully connected neural network with ≥2 hidden layers and 128-512 neurons per layer); The structured heterogeneous network (in a specific embodiment of the present invention, the structured heterogeneous network has 2 hidden layers, each with 64-128 neurons); The action policy network (in a specific embodiment of the present invention, the action policy network has two hidden layers, containing 256 and 128 neurons respectively). The input to the old execution network is defined as follows: it consists of state parameters and structural parameters, which are respectively represented by... and It means that among them They represent The position and pose of the manipulated object, the joint state of the dexterous hand, the position of the fingertips, the output values ​​of the tactile sensors, and other normalized values ​​of the environment and the object's state in the simulation will change as the action strategy is executed; and =[ ] respectively represent the first The length, width, and distance from the palmar plane of each finger of the candidate dexterous hand correspond to the normalized morphological and structural parameters that can be optimized for the dexterous hand, and also correspond to the set of dexterous hand morphological and structural parameters to be optimized in the open parameter automatic mapping module. First, the state parameters are input into the general policy network in the old execution network, and a feature vector for the current state is output. The structural parameters are input into the structural alienation network within the old execution network, outputting a feature vector for alienation control strategies based on different structural parameters. The two sets of vectors are merged into a single feature fusion vector through a concatenation operation. Then it is input into the action policy network, which outputs an action probability density distribution. ,in, express The action control variables output by the network are executed at all times. Finally, by expanding all subnetworks of the entire execution network, the complete conditional policy function can be obtained: , used to generate control signals for dexterous hand joints; S32: Constructing a value network with regularization constraints The value network employs an independently constructed multi-layer fully connected neural network; the value network Used to evaluate state value; for a value network, by... and These are inputs into the value network, which outputs a value assessment for the current state and an expected value for the future state. ,in, The mathematical expectation symbol is used to represent the expectation given the current state parameters. and structural parameters Under the given conditions, the expected value of the cumulative reward in the future after implementing the current strategy; This represents the reward discount factor, used to adjust the degree to which future rewards affect the current value estimate; Indicates from the current moment Backwards Instant rewards earned at each time step; This indicates that the future reward has been processed. Reward value after the discount; Accumulated returns for discounts, representing the period from time [time]. The beginning, the reward for every step in the future Multiply by discount factor Powers of and sums of; To address the issue of inconsistent value estimation under similar structural parameters, a structural consistency regularization term is introduced to calculate the total loss function of the value network, expressed as:

[0042] in, This represents the time steps sampled during training. Perform an empirical average, which means averaging the value estimation errors of all state samples within a training batch; That is, the expected value output of the value network; It is the value in the actual execution process; For structural consistency regularization terms The weighting coefficients are used to adjust the relative importance between the value fitting error and the structural continuity constraint; when When the value is large, the value network emphasizes the consistency of value output under similar structural parameters; when When the value network is smaller, it focuses more on fitting the value target obtained from actual sampling; It is a structural consistency regularization term, used to force similar structures to have similar value estimates; structural consistency regularization term The expression is:

[0043] in, Represents the state in the training samples and structural parameter combination index Empirical averaging is performed to statistically analyze the differences in the output of value functions corresponding to similar structural parameters under different states; To measure structure and Weights of similarity ; ; This represents the proximity threshold of structural parameters, used to determine the proximity of two sets of structural parameters. and Whether they belong to similar structures; only if the two structural parameters and The Euclidean distance is less than the threshold. The value difference between them is calculated only when they are in the same condition, which avoids forcibly flattening completely unrelated structures. The hyperparameters controlling the weight decay rate; two structural parameters and The closer the Euclidean distance, the better. The closer to 1, the greater the difference penalty; the greater the distance, The closer to 0, the smaller the penalty; The value network outputs the expected value of the input state. , used to calculate the advantage value of the network loss function; S33: Structure-aware generalized dominance estimation: Generalized dominance estimation based on structural perception: Calculate the required advantage function; where, Let be the structure-aware advantage function, i.e., the th The generalized dominance estimate under the structural parameters represents the value of the first-order dominance. Under certain structural parameters, at time... The portion of a value that is higher (or lower) than the average value of taking a certain action; It is the offset of the time step; subscript + Relative to the current time The future At that moment; It is a discount factor, which typically ranges from [0,1] and is used to reduce the weight of rewards or errors in future time steps; This is the GAE attenuation parameter (usually between 0 and 1), which controls the trade-off between bias and variance; For the current moment The future The temporal difference error at each moment represents the time difference in the structure. Next, the future moment + The "deviation between actual returns and expected value"; For structural weighting factors, According to the structure Timing difference error The variance is dynamically adjusted; when the variance of the value estimate of a certain structure is large (unstable), Reduced, leading to The overall value decreases, thus weakening the impact of long-term errors on the current advantage value (the strategy update becomes more conservative); conversely, the update becomes more aggressive. It is a very small smoothing constant (e.g., 10). −8 (), used to prevent the denominator from being zero; For the first Timing difference error under various structural parameters , It is a moment The reward value, It is an estimate of the value of the current state; Value estimation for the next state; After introducing the structure-aware mechanism, It will be based on the current structure The advantage value is adjusted based on the characteristics (such as mass and inertia) so that the policy update can more accurately reflect the physical characteristics of the current structure and avoid training oscillations caused by a "one-size-fits-all" advantage estimation. S34: Proximity Policy Optimization (PPO) Update Using the calculated advantage function The loss function of the execution network is constructed from the action probability distribution output by the policy network, and the parameters of the execution network are updated using a proximal policy optimization algorithm. This continues until the control strategy converges, thus obtaining the control strategy. and success rate ; The loss function expression for the execution network is:

[0044] in, The target loss function of the PPO-Clip algorithm; This represents the time steps sampled during training. Perform an empirical average, which means averaging the value estimation errors of all state samples within a training batch; For strategy ratio; It is the dominance function, which is estimated through the structure-aware generalized dominance method designed in this patent: The calculation yielded, where , It is the reward value at time t, and As structural weights, this design automatically reduces gradients with high structural uncertainty while making policy updates for stable structural parts stronger. Although the structure of the near-end policy optimization algorithm itself remains unchanged, the structure awareness capability is explicitly enhanced. This step innovatively constructs a separate execution network architecture and a structure-aware training mechanism to avoid gradient conflicts between different structures and improve training stability. Based on traditional generalized advantage estimation, it redefines the time-recursive form of the advantage function by introducing adaptive weight factors related to structural parameters. This allows the advantage function to decay over time while considering the uncertainty of value estimation under different structural conditions. This design makes policy updates more conservative when structural parameters change significantly or value estimation fluctuates strongly, thereby improving the stability and robustness of the policy optimization process.

[0045] A structural consistency regularization constraint mechanism is designed for the training process of value networks. This mechanism avoids non-physically meaningful jumps in value estimation in the structural parameter space by constraining the continuity of the value function output under similar structural parameter conditions. On the one hand, this regularization constraint significantly improves the generalization ability of the value network in the structural parameter dimension; on the other hand, it provides a reliable foundation for the stable calculation of the structure-aware advantage function, thereby enhancing the convergence and consistency of the structure-control joint optimization process as a whole. (4) Perform Bayesian optimization of outer loop structural parameters Step S4 aims to use a Bayesian optimization mechanism to achieve a converged control strategy. To evaluate the basis, an efficient search is performed in the global structural parameter space to find the structural parameter configuration with optimal performance; Step S4 includes: S41: The policy is obtained after the execution network trained on the dexterity task with the previous set of structural parameters converges. and success rate By randomizing the combination of parameters, This strategy is executed with n different sets of parameters each time, and the final success rate is obtained. ; S42: Use the parameter set and the corresponding success rate as the sampling set. A Gaussian process model is trained to fit the mapping relationship between structural parameters and performance; the expression of the Gaussian process model is: ,in, The function to be optimized is a black-box function. It is a structural parameter, a joint variable that contains all possible values ​​of the parameters; It is a mean function; For kernel functions; S43: After training the Gaussian process model, maximize the output of the Gaussian process model using the expectation boosting function as the acquisition function to obtain the new optimal observation structure parameters. ; S44: Invoke the open parameter automapping mechanism to generate the corresponding dexterous hand simulation model, input the new optimal observation structure parameters into the simulation environment, update the physical / simulation model, and use the converged control strategy. To obtain the corresponding true value, i.e., the true success rate. ; (5) Iteration control and termination judgment Determining the true success rate using a comparison function. Has the success rate reached the level corresponding to the previous set of structural parameters? :like Then use the open parameter auto-mapping module. This set of parameters updates the twin, returns to step S2 to update the dexterous hand model, and executes the closed-loop optimization system again; if Then and the corresponding actual success rate Add to the sampling set Update the Gaussian process model and repeat the above acquisition steps to obtain new candidate parameters, and then obtain the optimal observation value again. Repeat steps S41-S44 until the preset success rate threshold is reached, then the closed-loop optimization process stops, and the optimal parameter combination is obtained. The algorithm flow of the entire optimization strategy in this invention is as follows: First, initialize the dexterous hand baseline model and the initial twin generated by the open parameter automapping module, and set the iteration counter. and initial structural parameters and initialize the success rate. and empty sampling set ; Then, it enters the outer loop (when Execution occurs at specific times, with the inner loop being executed first in each iteration (executed when the execution is converged), and the structure parameters are then considered. Conditional sampling of action trajectories is used to calculate temporal difference errors and structural weights, thereby updating the value network. and execution network until the strategy converges and the success rate of the current structure is obtained. ; Next, we proceed to the outer ring structure optimization stage, where random sampling is performed. Group candidate structure parameters Calculate success rate through repetitive task scenarios And update the sampling set ; Subsequently, a Gaussian process (GP) surrogate model is trained to fit the mapping relationship between structural parameters and performance, and the optimal candidate parameters are selected through the acquisition function.

[0046] If new parameters If the corresponding prediction success rate is better than the current best value, then the twin is updated through the open parameter auto-mapping module, and the iteration counter is reset. And perform closed-loop optimization training again on the new structure; If the result is not better than the current best, the new sample is added to the sampling set and candidate parameters are reselected. Finally, when the number of iterations reaches the preset upper limit or the global success rate target is met, the algorithm outputs the optimal structure parameters. and their corresponding optimal control strategies To achieve coordinated optimization of the dexterous hand structure and control.

[0047] This invention constructs a low-cost, high-efficiency, and sustainably iterative dexterous hand structure-control joint optimization method by performing network structure decomposition, structure-aware advantage estimation, value function consistency regularization, and an embedded collaborative mechanism of Bayesian optimization and deep reinforcement learning. It effectively overcomes the problems of repetitive training, low sample efficiency, and discontinuous policies in existing technologies, and has significant technological advancements and engineering application value.

[0048] To further illustrate the technical solution of the present invention, the implementation process of the typical operation of "using a screwdriver to unscrew a screw with a three-finger dexterity hand" will be described in detail below.

[0049] For the screw-tightening task, a multi-objective reward function and a high-dimensional state input are designed; Reward function settings: To guide dexterous hands in performing fine motor skills, a reward function is used. It consists of six weighted factors: successful task completion (unscrewing) as a reward; the distance between the fingertip and the screwdriver as a reward; the contact area and force input received by the fingertip tactile sensor as a reward; the distance between the screwdriver and the screw head as a reward; dropping the screwdriver as a penalty; and the screwdriver failing to rotate after contacting the screw head as a penalty. Expression:

[0050] in, , , , , , These correspond to success rate, contact area, fingertip distance, contact force, drop penalty, and failure penalty, respectively. State space definition: The input to a general state network is defined as a normalized high-dimensional state vector. , It is composed of the dexterous hand's physical state, tactile perception information, and the pose information of the object being manipulated, and can be specifically broken down into:

[0051] in, It refers to the initial joint positions of a dexterous hand, i.e., the 11-dimensional joint angles. , Indicates the first The angles of each joint at the initial moment; Using 9-dimensional Cartesian coordinates, it refers to the apex spatial position. ,in, , indicating the first The Cartesian coordinates of a fingertip in three-dimensional space; The tactile sensor outputs a two-dimensional pressure value, which is a 12-dimensional tactile feature vector. Therefore, to facilitate its use as input to a neural network, this invention performs feature processing on the two-dimensional pressure distribution to construct a compact tactile feature vector: ,in, It is the average pressure value. That is the maximum pressure value. It is the contact area (defined as the percentage of pixels where the pressure exceeds a threshold). It is the pressure center; by piecing together the tactile features of all the fingertips, we obtain the overall tactile state vector: ; The three-dimensional positions of the screwdriver and screw. ; Euler angles for screwdrivers and screws, ; For structurally heterogeneous networks, the input is the normalized open-ended parameters. ; Inner loop execution network training: In this task, the inner loop performs network training, with a training scenario of 100 groups and a maximum training iteration count of 5000. The training termination condition is defined as more than 10 consecutive successful training iterations or reaching the maximum training iteration count. The threshold for a single successful training iteration is defined as at least 80 successful training iterations out of 100 groups. The network output convergence execution strategy is also defined. ; Outer ring Bayesian optimization: After the network training is completed in the inner loop, 60 different parameter combinations are obtained by randomizing the parameter combinations. Each of them Each is a parameter combination vector containing 23 optimizable parameters; Using an open-source automatic parameter mapping module, dexterous hands corresponding to all parameter combinations are constructed, generating 60 corresponding dexterous hand simulation models in batches; then, a convergent execution strategy is implemented. The system was deployed to these 60 models for testing, with each model executing 100 identical task scenarios to obtain the final success rate. ; Combine these 60 sets of parameters Corresponding success rate as the initial sampling set Through the initial sampling set Training a Gaussian process model: The mapping relationship between fitting parameters and success rate; sampling points Follows this distribution; After training the Gaussian process model, the acquisition function is used: The output of the Gaussian process model is collected to obtain the optimal observation value: ; Using this set of optimal observation parameters: A new dexterous hand is constructed using an open-ended parameter automapping module, and then the true value corresponding to the success rate is obtained through a convergent execution network. That is, the actual success rate; The comparison function determines whether the preset success rate threshold has been reached. The parameter set here is used for the previous set of parameters in the training strategy. Corresponding success rate If the threshold is reached Then, using this set of parameters, the twin is updated through the open parameter automapping module, and the inner loop network is trained again in a repetitive loop; if the target is not met... Then this set of parameters and the corresponding actual success rate Add to the sampling set Update the Gaussian process model to obtain the optimal observations again. ; After approximately 15 iterations of the inner and outer loops, the system finally outputs the optimal parameter combination. and corresponding strategies In the final test, the success rate of the dexterous hand under this configuration in the screw-tightening task remained stable at over 85%, which is 40 percentage points higher than the initial random structure (average success rate of about 45%), verifying the effectiveness of the collaborative optimization method of the present invention.

[0052] like Figure 4 As shown, a second aspect of the present invention provides a dexterous hand structural parameters and control strategy co-optimization system for implementing the above-mentioned co-optimization method, comprising: The benchmark model construction and parameterization module is used to construct a benchmark model of the dexterous hand and define its parameters. The symmetrical geometric structure model is used for each finger joint of the dexterous hand. The optimizable parameters are constrained into three types of core geometric parameters: joint length, width and spatial position of the joint relative to the palm plane, forming an open structural parameter vector. Each joint is designed as a detachable modular structure. The open parameter automatic mapping module is used to establish an open parameter automatic mapping mechanism. In response to the structural parameter vector to be updated output by the optimization algorithm, based on the pre-built parameter-model field mapping relationship, it traverses and updates the corresponding nodes in the dexterous hand structure description file, and performs parameterized intermediate representation conversion and parameter replacement on the geometric model file. It also updates the centroid position and inertial tensor derived dynamic parameters derived from geometric changes, and generates the updated dexterous hand simulation model. An embedded collaborative optimization closed-loop network module is used to perform reinforcement learning training of the inner-loop control strategy. The updated dexterous hand simulation model is imported into the training environment to construct an execution network including a general state network, a structural alienation network, and an action policy network. The general state network receives environmental and ontological state inputs, the structural alienation network receives the current structural parameter vector input, and the action policy network fuses the outputs of both to generate an action probability distribution. Combined with a value network with structural consistency regularization constraints and structural awareness generalized advantage estimation, the system is trained based on a proximal policy optimization algorithm until the task success rate under the current structure converges to a preset threshold. The Bayesian outer loop optimization module performs Bayesian optimization of the outer loop structure parameters. Based on the converged control strategy, it randomly samples multiple sets of candidate structure parameters and evaluates the corresponding task success rates. It constructs a sampling dataset containing structure parameters and success rates, trains a Gaussian process surrogate model, and selects the next optimal candidate structure parameters through a sampling function. It calls an open parameter automapping mechanism to generate the corresponding dexterous hand simulation model, inputs the candidate parameters into the simulation environment, and obtains the true success rate using the converged control strategy. The iterative control and termination judgment module is used to coordinate the interaction logic between the inner and outer loops and control the start and stop of the optimization process. If the actual success rate is better than the historical best value, the new parameters are used as the benchmark structural parameters, and the process returns to step S2 to update the dexterous hand model. Steps S3 to S4 are repeated. If the success rate is not better than the historical best value, the parameters and the corresponding success rate are added to the sampling dataset, the Gaussian process surrogate model is updated, and candidate parameters are reselected. The process is iteratively run until the global success rate target is reached, and the optimal structural parameters and the corresponding optimal control strategy are output. The optimal structural parameters can be mapped to the detachable joint module to realize the rapid conversion of simulation optimization results into a physical dexterous hand.

[0053] It should be noted that the dexterous hand structure parameter and control strategy co-optimization system provided in this embodiment can be a computer program (including program code) running on a computer device. For example, the dexterous hand structure parameter and control strategy co-optimization system is an application software. The dexterous hand structure parameter and control strategy co-optimization system can be used to execute the corresponding steps in the above-mentioned methods provided in the embodiments of this application.

[0054] This application also provides a computer-readable storage medium storing a computer program that is executed by a processor to implement... Figure 1 The methods provided in each step are detailed in the implementation methods provided in the above steps, and will not be repeated here.

[0055] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for co-optimizing structural parameters and control strategies of a dexterous hand, characterized in that, Includes the following steps: S1: Construct a baseline model of the dexterous hand and define its parameters; model each finger joint of the dexterous hand using a symmetrical geometric structure, and constrain the optimizable parameters into three core geometric parameters: joint length, width, and spatial position of the joint relative to the palmar plane, forming an open structural parameter vector; each joint is designed as a detachable modular structure. S2: Based on the pre-built parameter-model field mapping relationship, the corresponding nodes in the dexterous hand structure description file are traversed and updated respectively, and the geometric model file is transformed into a parameterized intermediate representation and replaced with parameters. The centroid position and inertial tensor derived dynamic parameters derived from geometric changes are updated synchronously to generate the updated dexterous hand simulation model. S3: Import the updated dexterous hand simulation model into the training environment, construct an execution network including a general state network, a structural alienation network and an action policy network, combine a value network with structural consistency regularization constraints and structural awareness generalized advantage estimation, and train based on the proximal policy optimization algorithm until the task success rate under the current structure converges to a preset threshold. S4: Randomly sample multiple sets of candidate structure parameters and evaluate the corresponding task success rates. Construct a sampled dataset containing structure parameters and success rates. Train a Gaussian process surrogate model and select the next optimal candidate structure parameter through a sampling function. Obtain the true success rate using a converged strategy. S5: If the actual success rate is better than the historical best value, then the new parameter is used as the baseline structural parameter, and the process returns to step S2 to update the dexterous hand model, and steps S3 to S4 are repeated; if it is not better, then the parameter and the corresponding success rate are added to the sampling dataset, the Gaussian process surrogate model is updated and candidate parameters are reselected, and the process is iterated until the global success rate target is reached. The optimal structural parameter and the corresponding optimal control strategy are output, and the optimal structural parameter is mapped to the detachable joint module to realize the rapid conversion of simulation optimization results into a physical dexterous hand.

2. The method for co-optimizing structural parameters and control strategies of a dexterous hand according to claim 1, characterized in that: The dexterous hand mentioned in step S1 includes the thumb, index finger, and middle finger, and adopts a hybrid drive method combining linear actuators and rotary actuators; The fingers of the dexterous hand are connected to the palm base through joints and are driven by multiple actuators to achieve multiple degrees of freedom. Each finger of the dexterous hand is equipped with an actuator embedded in the palm; the thumb achieves flexion movement at the proximal and distal ends of the palm through two independent actuators, the proximal ends of the index and middle fingers achieve flexion movement through linear actuators and transmission mechanisms, while the middle and distal ends of the palm achieve flexion control through independent rotary actuators.

3. The method for co-optimizing structural parameters and control strategies of a dexterous hand according to claim 2, characterized in that: The open structure parameter vector in step S1 has a dimension of 23, including the length and width of the proximal, middle, and distal palmar joints of the index and middle fingers, the length and width of the proximal and distal palmar joints of the thumb, and the spatial position parameters of the preset joints or joints of the index, middle, and thumb relative to the palmar plane.

4. A method for co-optimizing the structural parameters and control strategy of a dexterous hand according to any one of claims 1-3, characterized in that, Step S2 includes: The structural parameters in step S1 are preprocessed to establish a parameter-model field mapping table; the parameter-model field mapping table includes a structural description file field mapping table and a geometric model file field mapping table. After the optimization algorithm outputs the structural parameter vector to be updated, the system parses the original structural description file, constructs a node tree, and traverses and replaces the node fields corresponding to the target parameters according to the mapping table, only modifying the parts directly related to the optimization parameters, keeping the model topology and assembly relationship unchanged; For geometric model files, first convert them into an editable intermediate representation, then iterate through and update the relevant geometric feature dimensions according to the geometric parameter mapping table, and finally restore the modified intermediate representation file to a standard geometric model file. After the model is updated, the derived dynamic parameters, including joint mass, center of mass position, and inertia tensor, are automatically calculated and updated based on the changes in joint geometry.

5. A method for co-optimizing the structural parameters and control strategy of a dexterous hand according to any one of claims 1-3, characterized in that: In step S3, the general state network receives the environment and ontology state inputs, the structural alienation network receives the current structural parameter vector inputs, and the action strategy network fuses the outputs of both to generate an action probability distribution.

6. A method for co-optimizing the structural parameters and control strategy of a dexterous hand according to any one of claims 1-3, characterized in that: The total loss function of the value network described in step S3 is expressed as follows: in, This represents the time steps sampled during training. Perform an empirical average; That is, the expected value output of the value network; It is the value in the actual execution process; For structural consistency regularization terms Weighting coefficients; It is a structural consistency regularization term; Structural consistency regularization The expression is: in, Represents the state in the training samples and structural parameter combination index Empirical averaging is performed to statistically analyze the differences in the output of value functions corresponding to similar structural parameters under different states; This represents the proximity threshold of structural parameters, used to determine the proximity of two sets of structural parameters. and Do they belong to similar structures? To measure structure and Weights of similarity ; This is a hyperparameter used to control the rate of weight decay.

7. A method for co-optimizing the structural parameters and control strategy of a dexterous hand according to any one of claims 1-3 or 6, characterized in that: The expression for the structure-aware generalized advantage estimation described in step S3 is: in, For the first Generalized dominance estimates under structural parameters; It is the offset of the time step; subscript + Relative to the current time The future At that moment; It is a discount factor, and its value is usually between [0,1]. For GAE attenuation parameters; For the current moment The future The timing difference error at each moment; For structural weighting factors; The expression for the structure weight factor is: in, For the first Timing difference error under various structural parameters , It is a moment The reward value; It is an estimate of the value of the current state; This is a value estimate for the next state.

8. A method for co-optimizing the structural parameters and control strategy of a dexterous hand according to any one of claims 1-3 or 6, characterized in that: The loss function expression for the network executed in step S3 is: in, The target loss function of the PPO-Clip algorithm; This represents the time steps sampled during training. Perform an empirical average; For strategy ratio; It is the dominance function.

9. A method for co-optimizing the structural parameters and control strategy of a dexterous hand according to any one of claims 1-3 or 6, characterized in that: The Gaussian process model expression in step S4 is: ,in, The function to be optimized is a black-box function. These are structural parameters, containing all possible values ​​for all parameters; It is a mean function; For kernel functions; The acquisition function used in step S4 is the expected boost function.

10. A system for collaborative optimization of structural parameters and control strategies of a dexterous hand, characterized in that, A method for co-optimizing the structural parameters and control strategy of a dexterous hand as described in any one of claims 1-9, comprising: The benchmark model construction and parameterization module is used to construct a benchmark model of the dexterous hand and define its parameters. The symmetrical geometric structure model is used for each finger joint of the dexterous hand. The optimizable parameters are constrained into three types of core geometric parameters: joint length, width and spatial position of the joint relative to the palm plane, forming an open structural parameter vector. Each joint is designed as a detachable modular structure. The open parameter automatic mapping module is used to establish an open parameter automatic mapping mechanism. Based on the pre-built parameter-model field mapping relationship, it traverses and updates the corresponding nodes in the dexterous hand structure description file, and performs parameterized intermediate representation conversion and parameter replacement on the geometric model file. It also updates the centroid position and inertial tensor derived dynamic parameters derived from geometric changes, and generates the updated dexterous hand simulation model. An embedded collaborative optimization closed-loop network module is used to perform reinforcement learning training of the inner-loop control strategy. The updated dexterous hand simulation model is imported into the training environment to construct an execution network including a general state network, a structural alienation network, and an action policy network. The general state network receives environmental and ontological state inputs, the structural alienation network receives the current structural parameter vector input, and the action policy network fuses the outputs of both to generate an action probability distribution. Combined with a value network with structural consistency regularization constraints and structural awareness generalized advantage estimation, the system is trained based on a proximal policy optimization algorithm until the task success rate under the current structure converges to a preset threshold. The Bayesian outer loop optimization module performs Bayesian optimization of the outer loop structure parameters, randomly samples multiple sets of candidate structure parameters and evaluates the corresponding task success rates, constructs a sampling dataset containing structure parameters and success rates, trains a Gaussian process surrogate model, selects the next optimal candidate structure parameters through a sampling function, calls an open parameter automatic mapping mechanism to generate the corresponding dexterous hand simulation model, inputs the candidate parameters into the simulation environment, and obtains the true success rate using the converged control strategy. The iterative control and termination judgment module is used to coordinate the interaction logic between the inner and outer loops and control the start and stop of the optimization process. If the actual success rate is better than the historical best value, the new parameters are used as the benchmark structural parameters, and the process returns to step S2 to update the dexterous hand model. Steps S3 to S4 are repeated. If the success rate is not better than the historical best value, the parameters and the corresponding success rate are added to the sampling dataset, the Gaussian process surrogate model is updated, and candidate parameters are reselected. The process is iteratively run until the global success rate target is reached, and the optimal structural parameters and the corresponding optimal control strategy are output. The optimal structural parameters can be mapped to the detachable joint module to realize the rapid conversion of simulation optimization results into a physical dexterous hand.

Images