Double-layer optimization design method for dexterous hand directly driven by micro motor
By employing a dual-layer optimization design method for direct-drive dexterous hands using micro-motors, the problems of high hardware trial-and-error rates and operational deviations from requirements in existing designs are solved. This achieves efficient configuration optimization and enhanced safety, ensuring the physical feasibility and operational practicality of the design.
Patent Information
- Application Number
- CN202610300934.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-12
- Publication Date
- 2026-04-24
AI Technical Summary
Existing designs for direct-drive dexterous hands using micro motors lack systematic scientific guidance, resulting in high hardware trial-and-error costs, design schemes that deviate from real-world operational needs, and poor coordination between macroscopic topology and microscopic proportions, leading to frequent hardware failures.
A dual-layer optimization design method for direct-drive dexterous hands using micro-motors is adopted. Through configuration design and fine-tuning of proportional parameters, combined with Monte Carlo random sampling, covariance matrix adaptive evolution strategy and real human motion data, the kinematic topology distribution and proportional parameters of the finger-thumb are optimized, a parameterized model library is constructed, and a high-fidelity physics engine is verified.
Significantly reduces hardware trial-and-error costs, expands the collaborative workspace of dexterous hands, improves operational safety and practicality, and ensures the physical feasibility of the design and simplification of the mechanical structure.
Smart Images
Figure CN121912414A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of robot structure design and intelligent optimization technology, and specifically relates to a two-layer optimization design method for a micro motor direct-drive dexterous hand. Background Technology
[0002] As the core execution component of humanoid robots, the dexterous hand's morphological design, including the geometric distribution of the thumb, the length of the finger joints, the position of the finger base, and the integration method of the motors within the fingers, directly determines the dexterous hand's grasping ability and operational performance, making it a key research focus in the field of robotics.
[0003] Currently, dexterity hand design methods largely rely on engineers' experience and simple bionic principles, lacking systematic parameter exploration methods, making it difficult to guarantee the global optimality of the dexterity hand configuration. Existing dexterity hand designs are mainly divided into two categories. The first category is based on rope-driven or complex linkage mechanisms. Although this type of design can effectively reduce the mass of the fingertips, it requires complex mechanical structures such as gear sets and long transmission cables, which not only leads to a large overall size of the dexterity hand, but also generates nonlinear errors during control, affecting operational accuracy. The second category is the emerging micro-motor direct-drive dexterity hand design. By placing the micro-motor directly at the corresponding degree of freedom (DOF) position of the human hand joint, it achieves a highly compact dexterity hand structure and direct drive without backlash, solving some of the defects of rope-driven or linkage-based dexterity hands.
[0004] However, the design of this type of direct-drive dexterous hand with "direct placement of micro-motors" still lacks systematic scientific guidance and faces several technical challenges: First, there is a lack of systematic design methods for direct-drive motor structures. Existing dexterous hand morphological optimizations are mostly based on abstract kinematic line models, failing to consider the actual physical volume limitations imposed by placing micro-motors directly at the joints, and lacking a systematic configuration optimization scheme adapted to direct-drive micro-motors. Second, hardware trial and error and testing costs are extremely high. Manufacturing direct-drive dexterous hands relies on high-strength precision-machined parts such as CNC metal parts and high-strength 3D-printed parts. Relying solely on heuristic trial and error based on human intuition would lead to frequent hardware iterations, resulting in high manufacturing costs and... Third, there is a lack of systematic coordination between macroscopic topology and microscopic proportions. Existing optimization methods are often limited to single-detail optimization, such as optimizing only the length of the knuckles, ignoring the decisive influence of macroscopic topological structures such as the distribution pattern of the motors at the base of the thumb and the integrated arrangement of the finger motors on the collaborative workspace of dexterous hands. Fourth, the design scheme deviates from the actual operational needs and is prone to hardware damage. Many pure geometric optimization methods do not incorporate real human motion prior data. The optimized configuration theoretically has a large workspace, but when actually performing anthropomorphic tasks, it is very easy for the mechanical structure to have serious self-collision or exceed the actual physical limits of the motor, which in turn leads to overload of the direct drive motor and hardware damage.
[0005] Therefore, there is an urgent need to develop a configuration and proportional parameter optimization method specifically tailored for micro-motor direct-drive dexterous hands, which takes into account both physical feasibility and practical operational requirements, in order to overcome the aforementioned shortcomings of existing technologies. Summary of the Invention
[0006] To address the aforementioned problems, the purpose of this invention is to provide a dual-layer optimization design method for a micro-motor direct-drive dexterous hand. This method is specifically tailored for micro-motor direct-drive structures, achieving synergistic optimization of macroscopic topology and microscopic proportional parameters, effectively reducing hardware trial-and-error costs, and improving the collaborative workspace and operational safety of the dexterous hand.
[0007] The technical solution of this invention is: a dual-layer optimization design method for a micro-motor direct-drive dexterous hand. This method directly arranges the micro-motor at each degree of freedom joint of the dexterous hand without using complex mechanical transmission structures such as gear sets. It includes two stages: sequential configuration design and fine-tuning of proportional parameters. Specifically, the dual-layer optimization design method for a micro-motor direct-drive dexterous hand includes: S1. Extract the geometric dimensions of the dexterous hand shape as genetic variables, set the physical boundary constraints of the genetic variables, and construct a parametric model library that can be dynamically instantiated by the physics engine through normalization mapping combined with template rendering technology.
[0008] S2. For each hand configuration in the candidate topology library, Monte Carlo random sampling is performed in the joint space. The collision points of effective contact between the thumb and the fingertips of the other fingers are calculated through forward kinematics. The convex hull volume formed by the collision points of effective contact between the thumb and the fingertips of the other fingers is calculated. The optimal configuration with the largest cooperative workspace is selected by adopting the covariance matrix adaptive evolution strategy with the goal of maximizing the envelope volume.
[0009] S3. Based on the optimal configuration, a sequence of real human motion data containing complex movements is introduced. A two-layer optimization architecture is used to refine the micro-proportional parameters of the dexterous hand, outputting the final optimal physical size blueprint for the dexterous hand. The two-layer optimization architecture includes an outer framework and an inner framework. The outer framework uses the CMA-ES algorithm to generate and update candidate schemes in the morphological parameter space. The inner framework uses an inner evaluator to load the human motion data sequence and perform high-fidelity pose matching on the currently generated candidate physical models.
[0010] Furthermore, the gene values of the gene variables described in step S1 are converted into physical dimensions through normalization mapping, and then the physical dimensions are dynamically updated to the XML model of the simulation physics engine through Jinja2 template rendering technology.
[0011] The physical dimensions are converted using a conversion formula, which is: ,in, Indicates the extracted first Normalized gene values of each gene variable. Indicates the first The physical size values corresponding to each gene variable. Indicates the first The minimum physical size corresponding to each gene variable. Represented as the first The maximum physical size corresponding to each gene variable.
[0012] Furthermore, the specific steps of step S2 are as follows: S21. Define multiple dexterous hand topologies in the physics engine and build a library of candidate configurations.
[0013] S22. For each hand configuration in the candidate configuration library, Monte Carlo random sampling is performed in the joint space. The collision points where the thumb effectively contacts the fingertips of the other fingers are calculated using forward kinematics. The convex hull volume formed by the collision points of the thumb and the other fingertips is calculated. The sum of the convex hull volumes of the other fingers is used as the cooperative workspace volume. The cooperative workspace volume is determined according to the formula... The calculation yielded that, Indicates the volume of the collaborative workspace. It represents the collection of all fingers except the thumb. This represents the set of valid collision points for the corresponding finger.
[0014] S23. The Covariance Matrix Adaptive Evolution Strategy (CMA-ES) is adopted to maximize the volume of the collaborative workspace. The negative value of the collaborative workspace volume is used as the fitness. The CMA-ES algorithm is run independently for all configuration patterns in the library. The optimal configuration with the largest collaborative space and easy motor physical integration is selected by combining the evolution trend of the parameters.
[0015] Furthermore, the hand configuration of the candidate topology library mentioned in step S21 includes the thumb design and the motor distribution method.
[0016] The thumb design includes a first mode, a second mode, a third mode, and a fourth mode; the first mode is the Cross_Straight mode, where the first two degrees of freedom rotation axes rotate first Z then Y and the second link is parallel to the reference plane; the second mode is the Cross_Inclined mode, where Z then Y and the second link is tilted; the third mode is the Parallel_Straight mode, where X then Z and the second link is parallel to the reference plane; and the fourth mode is the Parallel_Inclined mode, where X then Z and the second link is tilted.
[0017] The motor distribution method includes a first distribution method and a second distribution method. The first distribution method is a distributed arrangement of motors distributed in multiple finger joints, and the second distribution method is an integrated arrangement of motors concentrated in the same finger joint.
[0018] Furthermore, the specific steps of step S3 are as follows: S31. Lock the optimal configuration selected in step S2, and define the core gene of the micro-proportion of the dexterous hand to be finely optimized. The core gene includes the base offset, the specific installation polar coordinates of the thumb, and the proportion of each fingertip.
[0019] S32. The outer optimizer generates and updates candidate design blueprints in the morphological parameter space based on the CMA-ES algorithm.
[0020] S33. The inner evaluator loads real human motion data sequences, performs high-fidelity pose matching for each candidate physical model generated by the outer layer, monitors the robot's collision depth and joint overruns throughout the dynamic tracking process, constructs a comprehensive loss function, and calculates the loss value.
[0021] S34. The outer optimizer iteratively updates the morphological scaling parameters based on the comprehensive loss function value returned by the inner evaluator until the early stopping mechanism is satisfied, and outputs the optimal physical size blueprint of the final direct-drive dexterous hand.
[0022] Furthermore, the comprehensive loss function described in step S33 The calculation formula is: ,in, , , Indicates the weighting coefficient. This represents the pose matching loss. Indicates collision penalty loss, This indicates the limit of joint punishment.
[0023] Furthermore, the pose matching loss is obtained by calculating the mean square error between the robot key vector and the scaled human action target vector, and the calculation formula is as follows: ,in, This represents the total number of frames in the human motion data sequence. This represents the number of key vectors of the robot. Indicates the first The first frame robot Key vectors Indicates the first The human action frame after scaling A target vector.
[0024] Furthermore, a collision penalty is incurred when the contact distance between any two non-adjacent geometries in the simulation engine is less than a safety threshold. The formula for calculating the collision penalty loss is as follows: ;in, This represents the total number of frames in the human motion data sequence. Indicates the first The set of non-adjacent geometric contacts in a frame. This indicates the preset safety threshold for contact with geometric objects. This represents the actual contact distance between non-adjacent geometric objects.
[0025] Furthermore, a secondary penalty is applied to poses that approach or exceed the physical limits of the motor to encourage them to move away from the joint boundaries and prevent the motor from stalling. The formula for calculating the joint limit penalty is as follows: ;in, This represents the total number of frames in the human motion data sequence. This indicates the total number of joints in a dexterous hand. Represents the joint penalty function. Indicates the first Frame Dexterity Hands The actual angle of each joint.
[0026] Furthermore, the joint penalty function is defined as follows: ;in, Indicates the first Each joint has a pre-defined safety boundary. Indicates the first The minimum physical limit angle of each joint. Indicates the first The maximum physical limit angle of each joint.
[0027] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention is specifically designed for direct drive systems where micro motors are placed directly at the joints, avoiding the control delays and nonlinear errors caused by traditional wire drives or linkages. At the same time, by strictly adhering to the physical boundaries of the motor and setting constraints at the underlying parameterization stage, the physical feasibility of the optimized solution is ensured.
[0028] This invention completely replaces the trial-and-error process that relies on human intuition and frequent 3D printing / CNC machining by using a two-stage framework of "first designing the configuration and then fine-tuning the scale parameters" and a high-fidelity physics engine for virtual prototype verification. This significantly reduces the trial-and-error cost of high-strength machined parts and greatly saves time and economic costs for hardware iteration.
[0029] This invention employs an innovative Monte Carlo co-collision envelope volume collaborative workspace evaluation method during the configuration design phase, which macroscopically locks in the optimal finger-thumb kinematic topology distribution pattern, significantly expanding the effective physical collaborative space of dexterous hands.
[0030] In the configuration design stage, this invention not only relies on the final envelope volume index to select the configuration, but also creatively extracts the natural evolution trend of morphological parameters during the operation of the evolutionary algorithm. This trend is used as a reliable basis for determining the physical arrangement of the motor. Under the premise of ensuring sufficient internal physical space, the mechanical complexity of the direct drive multi-degree-of-freedom structure is greatly simplified.
[0031] This invention introduces prior data of real human motion as a dynamic test boundary during the fine-tuning stage of the scaling parameters. Based on this, a composite comprehensive loss function is constructed, which includes pose matching loss, collision penalty loss, and joint limit penalty. This function can effectively eliminate scaling parameters that are "theoretically feasible but easily self-damaged under anthropomorphic operation". This provides a safety guarantee for high-intensity direct assembly from simulation to reality, and enhances the operational practicality of dexterous hands in real tasks and the extreme protection capability of hardware. Attached Figure Description
[0032] Figure 1 This is a process flow diagram of the present invention; Figure 2 This is a schematic diagram of the dexterous hand topology of four candidate thumb distribution patterns in the application examples of the present invention; Figure 3 This is a schematic diagram of the spatial-physical mapping relationship of gene parameters of two thumb distribution patterns, Cross-Inclined and Parallel-Inclined, in the dexterity hand abstract model, as an application example of this invention. Figure 4 This is a schematic diagram of two possible micro motor distribution methods in an application example of the present invention. Detailed Implementation
[0033] The following is combined Figures 1 to 4 The specific embodiments of the present invention will be described in detail below.
[0034] Example like Figure 1 The method shown is a two-layer optimization design method for a micro-motor direct-drive dexterous hand. This method places the micro-motor directly at each degree of freedom joint of the dexterous hand without using complex mechanical transmission structures such as gear sets. It includes two stages: configuration design and fine-tuning of proportional parameters. The specific components of the two-layer optimization design method for a micro-motor direct-drive dexterous hand are as follows: S1. Extract the geometric dimensions of the dexterous hand shape as genetic variables, set the physical boundary constraints of the genetic variables, and construct a parametric model library that can be dynamically instantiated by the physics engine through normalization mapping combined with template rendering technology.
[0035] S2. For each hand configuration in the candidate topology library, Monte Carlo random sampling is performed in the joint space. The collision points of effective contact between the thumb and the fingertips of the other fingers are calculated through forward kinematics. The convex hull volume formed by the collision points of effective contact between the thumb and the fingertips of the other fingers is calculated. The optimal configuration with the largest cooperative workspace is selected by adopting the covariance matrix adaptive evolution strategy with the goal of maximizing the envelope volume.
[0036] S3. Based on the optimal configuration, a sequence of real human motion data containing complex movements is introduced. A two-layer optimization architecture is used to refine the micro-proportional parameters of the dexterous hand, outputting the final optimal physical size blueprint for the dexterous hand. The two-layer optimization architecture includes an outer framework and an inner framework. The outer framework uses the CMA-ES algorithm to generate and update candidate schemes in the morphological parameter space. The inner framework uses an inner evaluator to load the human motion data sequence and perform high-fidelity pose matching on the currently generated candidate physical models.
[0037] Preferably, the gene values of the gene variables in step S1 are converted into physical dimensions through normalization mapping, and then the physical dimensions are dynamically updated to the XML model of the simulation physics engine through Jinja2 template rendering technology.
[0038] Physical dimensions are converted using a conversion formula, which is: ,in, Indicates the extracted first Normalized gene values of each gene variable. Indicates the first The physical size values corresponding to each gene variable. Indicates the first The minimum physical size corresponding to each gene variable. Represented as the first The maximum physical size corresponding to each gene variable.
[0039] Preferably, the specific steps of step S2 are as follows: S21. Define multiple dexterous hand topologies in the physics engine and build a library of candidate configurations.
[0040] S22. For each hand configuration in the candidate configuration library, perform Monte Carlo random sampling in the joint space. Calculate the collision points where the thumb effectively contacts the fingertips of the other fingers using forward kinematics. Calculate the convex hull volume formed by the collision points where the thumb effectively contacts the fingertips of the other fingers. Sum the convex hull volumes of the other fingers as the collaborative workspace volume. The collaborative workspace volume is determined according to the formula... The calculation yielded that, Indicates the volume of the collaborative workspace. It represents the collection of all fingers except the thumb. This represents the set of valid collision points for the corresponding finger.
[0041] S23. The Covariance Matrix Adaptive Evolution Strategy (CMA-ES) is adopted to maximize the volume of the collaborative workspace. The negative value of the collaborative workspace volume is used as the fitness. The CMA-ES algorithm is run independently for all configuration patterns in the library. The optimal configuration with the largest collaborative space and easy motor physical integration is selected by combining the evolution trend of the parameters.
[0042] Preferably, the hand configuration of the candidate topology library in step S21 includes the thumb design and the motor distribution method.
[0043] The thumb design includes four modes: mode 1, mode 2, mode 3, and mode 4. Mode 1 is the Cross_Straight mode, where the first two degrees of freedom rotation axes rotate Z first and then Y, and the second link is parallel to the reference plane. Mode 2 is the Cross_Inclined mode, where Z first and then Y, and the second link is tilted. Mode 3 is the Parallel_Straight mode, where X first and then Z, and the second link is parallel to the reference plane. Mode 4 is the Parallel_Inclined mode, where X first and then Z, and the second link is tilted.
[0044] The motor distribution methods include a first distribution method and a second distribution method. The first distribution method is a distributed arrangement in which the motors are distributed in multiple finger joints, and the second distribution method is an integrated arrangement in which the motors are centrally integrated in the same finger joint.
[0045] Preferably, the specific steps of step S3 are as follows: S31. Lock the optimal configuration selected in step S2, and define the core genes of the micro-proportion of the dexterous hand to be finely optimized. The core genes include the base offset, the specific installation polar coordinates of the thumb, and the proportions of each fingertip.
[0046] S32. The outer optimizer generates and updates candidate design blueprints in the morphological parameter space based on the CMA-ES algorithm.
[0047] S33. The inner evaluator loads real human motion data sequences, performs high-fidelity pose matching for each candidate physical model generated by the outer layer, monitors the robot's collision depth and joint overruns throughout the dynamic tracking process, constructs a comprehensive loss function, and calculates the loss value.
[0048] S34. The outer optimizer iteratively updates the morphological scaling parameters based on the comprehensive loss function value returned by the inner evaluator until the early stopping mechanism is satisfied, and outputs the optimal physical size blueprint of the final direct-drive dexterous hand.
[0049] Preferably, the comprehensive loss function in step S33 The calculation formula is: , , , Indicates the weighting coefficient. This represents the pose matching loss. Indicates collision penalty loss, This indicates the limit of joint punishment.
[0050] Preferably, the pose matching loss is obtained by calculating the mean square error between the robot key vector and the scaled human action target vector, and the calculation formula is as follows: ,in, This represents the total number of frames in the human motion data sequence. This represents the number of key vectors of the robot. Indicates the first The first frame robot Key vectors Indicates the first The human action frame after scaling A target vector.
[0051] Preferably, a collision penalty is generated when the contact distance between any two non-adjacent geometries in the simulation engine is less than a safety threshold. The formula for calculating the collision penalty loss is as follows: ;in, This represents the total number of frames in the human motion data sequence. Indicates the first The set of non-adjacent geometric contacts in a frame. This indicates the preset safety threshold for contact with geometric objects. This represents the actual contact distance between non-adjacent geometric objects.
[0052] Preferably, a secondary penalty is applied to poses that approach or exceed the physical limits of the motor to encourage them to move away from the joint boundaries and prevent the motor from stalling. The formula for calculating the joint limit penalty is: ;in, This represents the total number of frames in the human motion data sequence. This indicates the total number of joints in a dexterous hand. Represents the joint penalty function. Indicates the first Frame Dexterity Hands The actual angle of each joint.
[0053] Preferably, the joint penalty function is defined as follows: ;in, Indicates the first Each joint has a pre-defined safety boundary. Indicates the first The minimum physical limit angle of each joint. Indicates the first The maximum physical limit angle of each joint.
[0054] It should be noted that the CMA-ES algorithm in this embodiment can be replaced by other derivative-free black-box optimization algorithms such as Genetic Algorithm (GA) and Particle Swarm Optimization (PSO). The pose matching algorithm in the inner evaluator can be replaced by other inverse kinematics solvers. The optimization effect of this invention can still be achieved after the replacement, and all of these are within the protection scope of this invention.
[0055] Application examples This application example is based on the dual-layer optimization design method for micro-motor direct-drive dexterous hands proposed in the embodiments. It uses a hollow cup micro servo motor as the direct drive power source and MuJoCo as the simulation physics engine. It elaborates on the dual-layer optimization method for configuration design and scale parameter fine-tuning for micro-motor direct-drive dexterous hands. The complete process of this application example is divided into three parts: parametric modeling, configuration design, and scale parameter fine-tuning. Among them, configuration design and scale parameter fine-tuning are the two core optimization stages.
[0056] Preliminary steps: Unified construction of gene parameter space and establishment of parameterized model library The size variables of the micro motor direct-drive hand are extracted as genetic variables. The size variables include finger base offset, length of each finger joint, thumb polar coordinate parameters, finger base spatial offset, and fingertip physical length. Based on the physical length and width data of the hollow cup micro servo motor, the lower limit constraint of the genetic variables is set to prevent motor interference, and the upper limit constraint of the variables is set to prevent the dexterous hand from becoming too large.
[0057] Gene values of gene variables The normalized mapping is used to convert the data to physical dimensions. The conversion formula is as follows: ,in, Indicates the extracted first Normalized gene values of each gene variable. Indicates the first The physical size values corresponding to each gene variable. Indicates the first The minimum physical size corresponding to each gene variable. Represented as the first The maximum physical dimensions corresponding to each gene variable are used. Jinja2 template rendering technology is used to dynamically update the physical dimensions corresponding to each gene variable after conversion to the XML model of the MuJoCo simulation physics engine. A parameterized model library that can be dynamically instantiated by the MuJoCo physics engine is constructed to ensure that all subsequent optimization steps are performed under the strict constraints of the real motor physical envelope.
[0058] Phase 1: Configuration design based on Monte Carlo collision envelope volume S21. Define the library of candidate configurations. In the MuJoCo physics engine, multiple dexterous hand topologies are defined, encompassing four thumb spatial distribution modes (Cross_Straight, Cross_Inclined, Parallel_Straight, Parallel_Inclined) and two finger motor integration modes (distributed arrangement, integrated arrangement), to construct a candidate configuration library. The thumb spatial distribution mode refers to the arrangement mode controlling the two degrees of freedom (CMC) of the thumb wrist-palmar joint. The four modes are: Cross_Straight mode (Z-first then Y, with the second link parallel to the reference plane); and Cross_... The four configurations are: Inclined, Parallel_Straight (X-first, Z-second, with the second link parallel to the reference plane), and Inclined (X-first, Z-second, with the second link tilted). The gene sequence dimensions and hardware physical limitations for each configuration are set according to actual design requirements. The total gene dimension is 12 for Parallel_Straight, 13 for Parallel_Inclined, 13 for Cross_Straight, and 14 for Cross_Inclined. The global gene parameter distribution for each of the four configurations is shown in Table 1.
[0059] Table 1: List of gene parameters under four modes like Figure 2 The diagram shows the dexterous hand topology for four possible thumb spatial distribution patterns. The diagram illustrates two motor arrangements, Cross and Parallel, and two linkage geometry angles, Straight and Inclined. All four patterns are arrangement patterns that control the two degrees of freedom of the thumb wrist-palm joint (CMC).
[0060] like Figure 3 As shown, this example illustrates the spatial-physical mapping relationship between the gene parameters of the two thumb distribution patterns, Cross-Inclined and Parallel-Inclined, on the dexterity hand abstract model. In the figure, Cross-I represents Cross-Inclined, and Parallel-I represents Parallel-Inclined.
[0061] like Figure 4 As shown, this application example has two distribution methods: the first distribution method is a distributed arrangement of motors distributed across multiple finger joints, and the second distribution method is an integrated arrangement of motors concentrated in the same finger joint.
[0062] S22. Effective collision point sampling based on Monte Carlo method For each hand configuration in the candidate configuration library, an XML model with corresponding parameters is generated in the simulation environment. N=10000 sets of joint angles are randomly generated in the joint space. The effective contact points between the thumb and other fingertips are calculated using forward kinematics. If the capsule bodies of the two fingertips collide and the contact point is located within the 10% region of the fingertips' ends, the coordinates of this point cloud are recorded. The volume envelope of the effective collision point cloud for each finger is calculated using a 3D convex hull algorithm. The sum of the convex hull volumes of the thumb and other fingers is used as the volume of the collaborative workspace. The calculation formula is: The calculation yielded that, Indicates the volume of the collaborative workspace. It represents the set of fingers other than the thumb. This represents the set of valid collision points for the corresponding finger.
[0063] S23. Execute parallel CMA-ES configuration screening and motor layout guidance: Adopt the covariance matrix adaptive evolution strategy CMA-ES, with the optimization objective of maximizing the collaborative workspace volume, and use the negative value of the collaborative workspace volume as the fitness. ,in, To represent fitness, the CMA-ES algorithm is run independently on all configuration patterns in the library. The maximum cooperative envelope volume after convergence is compared, and the optimal configuration with the largest cooperative space and easy physical integration of motors is selected based on the evolution trend of the parameters. In this embodiment, the optimal configuration selected is the Parallel_Inclined mode. At the same time, by analyzing the evolution trend of the proportion parameters of each phalanx in the evolution algorithm, for example, it is observed that the length of the middle phalanx shows a significant elongation trend during the evolution process, which provides a sufficient internal space basis for the internal integration of multiple proximal and distal interphalangeal joint micro motors. Using this evolution trend as a physical guideline, the integrated arrangement scheme of the finger motors is finalized, which greatly reduces mechanical complexity while maintaining performance. Then, this configuration architecture and motor arrangement scheme are brought into the second stage.
[0064] Phase Two: Fine-tuning of the two-layer scaling parameters based on human motion data S31. Establishing a fixed framework and fine-tuning genes The optimal configuration architecture and motor integration layout of Parallel_Inclined selected in the first stage are identified, and the core genes to be finely optimized are defined. The core genes include base offset, specific polar coordinates of thumb installation, and proportions of each fingertip.
[0065] S32, Introducing Real Human Motion Priors Load a sequence of complex human hand movements captured visually as a dynamic test boundary.
[0066] S33, Two-layer dynamic evaluation The outer optimizer generates and updates candidate design blueprints in the morphological parameter space based on the CMA-ES algorithm; the inner evaluator solves the inverse kinematics for each robot physical model generated by the outer layer frame by frame to achieve pose matching of human action sequences, and monitors the robot's collision depth and joint overruns throughout the dynamic tracking process, constructs a comprehensive loss function and calculates the loss value. Comprehensive loss function The calculation formula is: ,in, , , Indicates the weighting coefficient. This represents the pose matching loss. Indicates collision penalty loss, This indicates the limit of joint punishment.
[0067] The formula for calculating pose matching loss is: ,in, This represents the total number of frames in the human motion data sequence. This represents the number of key vectors of the robot. Indicates the first The first frame robot Key vectors Indicates the first The human action frame after scaling A target vector.
[0068] The formula for calculating collision penalty loss is: ;in, This represents the total number of frames in the human motion data sequence. Indicates the first The set of non-adjacent geometric contacts in a frame. This indicates the preset safety threshold for contact with geometric objects, as shown in this application example. The physical dimensions of the coreless motor are set at 0.005m. This represents the actual contact distance between non-adjacent geometric objects.
[0069] The formula for calculating the joint limit penalty is: ;in, This represents the total number of frames in the human motion data sequence. This indicates the total number of joints in a dexterous hand. Represents the joint penalty function. Indicates the first Frame Dexterity Hands The actual angles of each joint. Among them, the penalty function... Apply secondary penalties to poses that approach or exceed the physical limits of the motor, joint security boundary Set to 10% of the joint's physical limit range.
[0070] S34, Outer Layer Evolution Iteration The outer optimizer iteratively updates the morphological scaling parameters based on the comprehensive loss function value returned by the inner evaluator. When the early stopping mechanism is satisfied, the iteration stops, and the final direct-drive dexterous hand hardware blueprint containing the optimal microscale dimensions is output.
[0071] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and do not limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the protection scope of the present invention.
Claims
1. A dual-layer optimized design method for a micro-motor direct-drive dexterous hand, characterized in that, include: S1. Extract the geometric dimensions that control the shape of the dexterous hand as genetic variables, set the physical boundary constraints of the genetic variables, and construct a parametric model library that can be dynamically instantiated by the physics engine through normalization mapping combined with template rendering technology. S2. For each hand configuration in the candidate topology library, Monte Carlo random sampling is performed in the joint space. The collision points of effective contact between the thumb and the fingertips of the other fingers are calculated through forward kinematics. The convex hull volume formed by the collision points of effective contact between the thumb and the fingertips of the other fingers is calculated. The optimal configuration with the largest cooperative workspace is selected by adopting the covariance matrix adaptive evolution strategy with the goal of maximizing the envelope volume. S3. Based on the optimal configuration, introduce a sequence of real human motion data containing complex movements, refine the micro-proportional parameters of the dexterous hand through a two-layer optimization architecture, and output the optimal physical size blueprint of the dexterous hand. The two-layer optimization architecture includes an outer layer architecture and an inner layer architecture. The outer layer architecture uses the CMA-ES algorithm to generate and update candidate schemes in the morphological parameter space. The inner layer architecture uses an inner evaluator to load human motion data sequences and perform high-fidelity pose matching on the currently generated candidate physical models.
2. The dual-layer optimization design method for a micro-motor direct-drive dexterous hand as described in claim 1, characterized in that, The gene values of the gene variables described in step S1 are converted into physical dimensions through normalization mapping, and then the physical dimensions are dynamically updated to the XML model of the simulation physics engine through Jinja2 template rendering technology. The physical dimensions are converted using a conversion formula, which is: ,in, Indicates the extracted first Normalized gene values of each gene variable. Indicates the first The physical size values corresponding to each gene variable. Indicates the first The minimum physical size corresponding to each gene variable. Represented as the first The maximum physical size corresponding to each gene variable.
3. The dual-layer optimization design method for a micro-motor direct-drive dexterous hand as described in claim 1, characterized in that, The specific steps of step S2 are as follows: S21. Define multiple sets of dexterous hand topologies in the physics engine and build a library of candidate configurations; S22. For each hand configuration in the candidate configuration library, Monte Carlo random sampling is performed in the joint space. The collision points where the thumb effectively contacts the fingertips of the other fingers are calculated using forward kinematics. The convex hull volume formed by the collision points of the thumb and the other fingertips is calculated. The sum of the convex hull volumes of the other fingers is used as the cooperative workspace volume. The cooperative workspace volume is determined according to the formula... The calculation yielded that, Indicates the volume of the collaborative workspace. The group of fingers excluding the thumb. This represents the set of valid collision points for the corresponding finger. S23. The Covariance Matrix Adaptive Evolution Strategy (CMA-ES) is adopted to maximize the volume of the collaborative workspace. The negative value of the collaborative workspace volume is used as the fitness. The CMA-ES algorithm is run independently for all configuration patterns in the library. The optimal configuration with the largest collaborative space and easy motor physical integration is selected by combining the evolution trend of the parameters.
4. The dual-layer optimization design method for a micro-motor direct-drive dexterous hand as described in claim 3, characterized in that, The hand configuration of the candidate topology library mentioned in step S21 includes the thumb design and the motor distribution method; The thumb design includes a first mode, a second mode, a third mode, and a fourth mode; the first mode is a Cross_Straight mode where the first two degrees of freedom rotation axes rotate first Z then Y and the second link is parallel to the reference plane; the second mode is a Cross_Inclined mode where Z rotates first then Y and the second link is tilted; the third mode is a Parallel_Straight mode where X rotates first then Z and the second link is parallel to the reference plane; the fourth mode is a Parallel_Inclined mode where X rotates first then Z and the second link is tilted. The motor distribution method includes a first distribution method and a second distribution method. The first distribution method is a distributed arrangement of motors distributed in multiple finger joints, and the second distribution method is an integrated arrangement of motors concentrated in the same finger joint.
5. The dual-layer optimization design method for a micro-motor direct-drive dexterous hand as described in claim 1, characterized in that, The specific steps of step S3 are as follows: S31. Lock the optimal configuration selected in step S2, and define the core gene of the micro-proportion of the dexterous hand to be finely optimized. The core gene includes the base offset, the specific installation polar coordinates of the thumb, and the proportion of each fingertip. S32. The outer optimizer generates and updates candidate design blueprints in the morphological parameter space based on the CMA-ES algorithm. S33. The inner evaluator loads real human motion data sequences, performs high-fidelity pose matching for each candidate physical model generated by the outer layer, monitors the robot's collision depth and joint limit exceedance throughout the dynamic tracking process, constructs a comprehensive loss function and calculates the loss value. S34. The outer optimizer iteratively updates the morphological scaling parameters based on the comprehensive loss function value returned by the inner evaluator until the early stopping mechanism is satisfied, and outputs the optimal physical size blueprint of the final direct-drive dexterous hand.
6. The dual-layer optimization design method for a micro-motor direct-drive dexterous hand as described in claim 5, characterized in that, The comprehensive loss function described in step S33 The calculation formula is: ,in, , , Indicates the weighting coefficient. This represents the pose matching loss. Indicates collision penalty loss, This indicates the limit of joint punishment.
7. The dual-layer optimization design method for a micro-motor direct-drive dexterous hand as described in claim 6, characterized in that, The pose matching loss is obtained by calculating the mean square error between the robot key vector and the scaled human motion target vector, and the calculation formula is as follows: ,in, This represents the total number of frames in the human motion data sequence. This represents the number of key vectors of the robot. Indicates the first The first frame robot Key vectors Indicates the first The human action frame after scaling A target vector.
8. The dual-layer optimization design method for a micro-motor direct-drive dexterous hand as described in claim 6, characterized in that, A collision penalty occurs when the contact distance between any two non-adjacent geometries in the simulation engine is less than a safety threshold. The formula for calculating the collision penalty loss is as follows: ;in, This represents the total number of frames in the human motion data sequence. Indicates the first The set of non-adjacent geometric contacts in a frame. This indicates the preset safety threshold for contact with geometric objects. This represents the actual contact distance between non-adjacent geometric objects.
9. The dual-layer optimization design method for a micro-motor direct-drive dexterous hand as described in claim 6, characterized in that, Apply secondary penalties to poses that approach or exceed the physical limits of the motor to encourage them to move away from the joint boundaries and prevent the motor from stalling. The formula for calculating the joint limit penalty is as follows: ;in, This represents the total number of frames in the human motion data sequence. This indicates the total number of joints in a dexterous hand. Represents the joint penalty function. Indicates the first Frame Dexterity Hands The actual angle of each joint.
10. The dual-layer optimization design method for a micro-motor direct-drive dexterous hand as described in claim 9, characterized in that, The joint penalty function is defined as follows: ;in, Indicates the first Each joint has a pre-defined safety boundary. Indicates the first The minimum physical limit angle of each joint. Indicates the first The maximum physical limit angle of each joint.